Merge pull request #506 from microsoft/update-translations
🌐 Update translations via Co-op Translator
This commit is contained in:
Lee Stott
GitHub
commit
1f18f7a6cd
|
|
@ -1,8 +1,8 @@
|
|||
<!--
|
||||
CO_OP_TRANSLATOR_METADATA:
|
||||
{
|
||||
"original_hash": "f3a6b0ddf7e6e3f33b2a543baf086dc9",
|
||||
"translation_date": "2025-08-24T21:40:38+00:00",
|
||||
"original_hash": "07191303b7ea2aff1d47e2b0fe4bb862",
|
||||
"translation_date": "2025-08-31T09:19:33+00:00",
|
||||
"source_file": "README.md",
|
||||
"language_code": "hk"
|
||||
}
|
||||
|
|
@ -21,136 +21,147 @@ CO_OP_TRANSLATOR_METADATA:
|
|||
|
||||
[](https://discord.gg/zxKYvhSnVp?WT.mc_id=academic-000002-leestott)
|
||||
|
||||
# 初學者人工智能課程 - 課程大綱
|
||||
# 人工智能初學者課程 - 課程大綱
|
||||
|
||||
| 繪製的手繪筆記 ](./lessons/sketchnotes/ai-overview.png)|
|
||||
||
|
||||
|:---:|
|
||||
| 初學者人工智能 - _手繪筆記由 [@girlie_mac](https://twitter.com/girlie_mac) 繪製_ |
|
||||
| 人工智能初學者課程 - _速寫筆記由 [@girlie_mac](https://twitter.com/girlie_mac) 繪製_ |
|
||||
|
||||
透過我們為期 12 週、共 24 節課的課程，探索**人工智能**（AI）的世界！課程包括實用的教學、測驗和實驗室練習。這是一個適合初學者的課程，涵蓋了 TensorFlow 和 PyTorch 等工具，以及 AI 的倫理問題。
|
||||
探索 **人工智能** (AI) 的世界，通過我們為期 12 週、共 24 節課的課程！課程包括實用教學、測驗和實驗室練習。這是一個適合初學者的課程，涵蓋了如 TensorFlow 和 PyTorch 等工具，以及 AI 的倫理問題。
|
||||
|
||||
## 你將學到什麼
|
||||
### 🌐 多語言支持
|
||||
|
||||
#### 通過 GitHub Action 支持（自動化且始終保持最新）
|
||||
|
||||
[法文](../fr/README.md) | [西班牙文](../es/README.md) | [德文](../de/README.md) | [俄文](../ru/README.md) | [阿拉伯文](../ar/README.md) | [波斯文 (法爾西)](../fa/README.md) | [烏爾都文](../ur/README.md) | [中文 (簡體)](../zh/README.md) | [中文 (繁體，澳門)](../mo/README.md) | [中文 (繁體，香港)](./README.md) | [中文 (繁體，台灣)](../tw/README.md) | [日文](../ja/README.md) | [韓文](../ko/README.md) | [印地文](../hi/README.md) | [孟加拉文](../bn/README.md) | [馬拉地文](../mr/README.md) | [尼泊爾文](../ne/README.md) | [旁遮普文 (古木基)](../pa/README.md) | [葡萄牙文 (葡萄牙)](../pt/README.md) | [葡萄牙文 (巴西)](../br/README.md) | [意大利文](../it/README.md) | [波蘭文](../pl/README.md) | [土耳其文](../tr/README.md) | [希臘文](../el/README.md) | [泰文](../th/README.md) | [瑞典文](../sv/README.md) | [丹麥文](../da/README.md) | [挪威文](../no/README.md) | [芬蘭文](../fi/README.md) | [荷蘭文](../nl/README.md) | [希伯來文](../he/README.md) | [越南文](../vi/README.md) | [印尼文](../id/README.md) | [馬來文](../ms/README.md) | [他加祿文 (菲律賓)](../tl/README.md) | [斯瓦希里文](../sw/README.md) | [匈牙利文](../hu/README.md) | [捷克文](../cs/README.md) | [斯洛伐克文](../sk/README.md) | [羅馬尼亞文](../ro/README.md) | [保加利亞文](../bg/README.md) | [塞爾維亞文 (西里爾字母)](../sr/README.md) | [克羅地亞文](../hr/README.md) | [斯洛文尼亞文](../sl/README.md) | [烏克蘭文](../uk/README.md) | [緬甸文 (緬甸)](../my/README.md)
|
||||
|
||||
**如果您希望支持其他翻譯語言，請參考 [此處](https://github.com/Azure/co-op-translator/blob/main/getting_started/supported-languages.md)**
|
||||
|
||||
## 加入社群
|
||||
[](https://discord.gg/kzRShWzttr)
|
||||
|
||||
## 您將學到什麼
|
||||
|
||||
**[課程思維導圖](http://soshnikov.com/courses/ai-for-beginners/mindmap.html)**
|
||||
|
||||
在這個課程中，你將學到：
|
||||
在這個課程中，您將學到：
|
||||
|
||||
* 不同的人工智能方法，包括傳統的符號方法，使用**知識表示**和推理（[GOFAI](https://en.wikipedia.org/wiki/Symbolic_artificial_intelligence)）。
|
||||
* **神經網絡**和**深度學習**，這是現代 AI 的核心。我們將使用兩個最受歡迎的框架 [TensorFlow](http://Tensorflow.org) 和 [PyTorch](http://pytorch.org) 的代碼來說明這些重要主題背後的概念。
|
||||
* 用於處理圖像和文本的**神經架構**。我們將涵蓋一些近期的模型，但可能不包括最前沿的技術。
|
||||
* 不太流行的 AI 方法，例如**遺傳算法**和**多代理系統**。
|
||||
* 不同的人工智能方法，包括使用 **知識表示** 和推理的「經典」符號方法 ([GOFAI](https://en.wikipedia.org/wiki/Symbolic_artificial_intelligence))。
|
||||
* **神經網絡** 和 **深度學習**，這是現代 AI 的核心。我們將使用兩個最受歡迎的框架 [TensorFlow](http://Tensorflow.org) 和 [PyTorch](http://pytorch.org) 的代碼來說明這些重要主題背後的概念。
|
||||
* 用於處理圖像和文本的 **神經架構**。我們將涵蓋一些近期的模型，但可能不包括最前沿的技術。
|
||||
* 不太流行的 AI 方法，例如 **遺傳算法** 和 **多代理系統**。
|
||||
|
||||
我們不會在這個課程中涵蓋：
|
||||
本課程不會涵蓋：
|
||||
|
||||
> [在 Microsoft Learn 集合中找到本課程的所有額外資源](https://learn.microsoft.com/en-us/collections/7w28iy2xrqzdj0?WT.mc_id=academic-77998-bethanycheum)
|
||||
|
||||
* **AI 在商業中的應用案例**。建議參加 [Introduction to AI for business users](https://docs.microsoft.com/learn/paths/introduction-ai-for-business-users/?WT.mc_id=academic-77998-bethanycheum) 的學習路徑，或 [AI Business School](https://www.microsoft.com/ai/ai-business-school/?WT.mc_id=academic-77998-bethanycheum)，這是與 [INSEAD](https://www.insead.edu/) 合作開發的。
|
||||
* **經典機器學習**，這在我們的 [Machine Learning for Beginners Curriculum](http://github.com/Microsoft/ML-for-Beginners) 中有詳細描述。
|
||||
* 使用 **[Cognitive Services](https://azure.microsoft.com/services/cognitive-services/?WT.mc_id=academic-77998-bethanycheum)** 構建的實際 AI 應用。建議從 Microsoft Learn 的模組開始，例如 [vision](https://docs.microsoft.com/learn/paths/create-computer-vision-solutions-azure-cognitive-services/?WT.mc_id=academic-77998-bethanycheum)、[natural language processing](https://docs.microsoft.com/learn/paths/explore-natural-language-processing/?WT.mc_id=academic-77998-bethanycheum)、**[Generative AI with Azure OpenAI Service](https://learn.microsoft.com/en-us/training/paths/develop-ai-solutions-azure-openai/?WT.mc_id=academic-77998-bethanycheum)** 等。
|
||||
* 特定的 ML **雲框架**，例如 [Azure Machine Learning](https://azure.microsoft.com/services/machine-learning/?WT.mc_id=academic-77998-bethanycheum)、[Microsoft Fabric](https://learn.microsoft.com/en-us/training/paths/get-started-fabric/?WT.mc_id=academic-77998-bethanycheum) 或 [Azure Databricks](https://docs.microsoft.com/learn/paths/data-engineer-azure-databricks?WT.mc_id=academic-77998-bethanycheum)。建議參考 [Build and operate machine learning solutions with Azure Machine Learning](https://docs.microsoft.com/learn/paths/build-ai-solutions-with-azure-ml-service/?WT.mc_id=academic-77998-bethanycheum) 和 [Build and Operate Machine Learning Solutions with Azure Databricks](https://docs.microsoft.com/learn/paths/build-operate-machine-learning-solutions-azure-databricks/?WT.mc_id=academic-77998-bethanycheum) 的學習路徑。
|
||||
* **對話式 AI** 和 **聊天機器人**。有一個單獨的 [Create conversational AI solutions](https://docs.microsoft.com/learn/paths/create-conversational-ai-solutions/?WT.mc_id=academic-77998-bethanycheum) 學習路徑，你也可以參考 [這篇博客文章](https://soshnikov.com/azure/hello-bot-conversational-ai-on-microsoft-platform/) 獲取更多細節。
|
||||
* **深度學習的數學基礎**。建議參考 Ian Goodfellow、Yoshua Bengio 和 Aaron Courville 的 [Deep Learning](https://www.amazon.com/Deep-Learning-Adaptive-Computation-Machine/dp/0262035618)，該書也可以在線獲取：[https://www.deeplearningbook.org/](https://www.deeplearningbook.org/)。
|
||||
* **AI 在商業中的應用案例**。建議參加 [商業用戶的 AI 入門](https://docs.microsoft.com/learn/paths/introduction-ai-for-business-users/?WT.mc_id=academic-77998-bethanycheum) 學習路徑，或 [AI 商業學院](https://www.microsoft.com/ai/ai-business-school/?WT.mc_id=academic-77998-bethanycheum)，該課程與 [INSEAD](https://www.insead.edu/) 合作開發。
|
||||
* **經典機器學習**，這在我們的 [機器學習初學者課程](http://github.com/Microsoft/ML-for-Beginners) 中有詳細描述。
|
||||
* 使用 **[認知服務](https://azure.microsoft.com/services/cognitive-services/?WT.mc_id=academic-77998-bethanycheum)** 構建的實際 AI 應用。建議從 Microsoft Learn 的模組開始學習 [視覺](https://docs.microsoft.com/learn/paths/create-computer-vision-solutions-azure-cognitive-services/?WT.mc_id=academic-77998-bethanycheum)、[自然語言處理](https://docs.microsoft.com/learn/paths/explore-natural-language-processing/?WT.mc_id=academic-77998-bethanycheum)、**[Azure OpenAI Service 的生成式 AI](https://learn.microsoft.com/en-us/training/paths/develop-ai-solutions-azure-openai/?WT.mc_id=academic-77998-bethanycheum)** 等內容。
|
||||
* 特定的 ML **雲框架**，如 [Azure Machine Learning](https://azure.microsoft.com/services/machine-learning/?WT.mc_id=academic-77998-bethanycheum)、[Microsoft Fabric](https://learn.microsoft.com/en-us/training/paths/get-started-fabric/?WT.mc_id=academic-77998-bethanycheum) 或 [Azure Databricks](https://docs.microsoft.com/learn/paths/data-engineer-azure-databricks?WT.mc_id=academic-77998-bethanycheum)。建議參考 [使用 Azure Machine Learning 構建和運營機器學習解決方案](https://docs.microsoft.com/learn/paths/build-ai-solutions-with-azure-ml-service/?WT.mc_id=academic-77998-bethanycheum) 和 [使用 Azure Databricks 構建和運營機器學習解決方案](https://docs.microsoft.com/learn/paths/build-operate-machine-learning-solutions-azure-databricks/?WT.mc_id=academic-77998-bethanycheum) 學習路徑。
|
||||
* **對話式 AI** 和 **聊天機器人**。有一個單獨的 [創建對話式 AI 解決方案](https://docs.microsoft.com/learn/paths/create-conversational-ai-solutions/?WT.mc_id=academic-77998-bethanycheum) 學習路徑，您也可以參考 [這篇博客文章](https://soshnikov.com/azure/hello-bot-conversational-ai-on-microsoft-platform/) 獲取更多細節。
|
||||
* **深度學習的數學基礎**。建議參考 Ian Goodfellow、Yoshua Bengio 和 Aaron Courville 的 [Deep Learning](https://www.amazon.com/Deep-Learning-Adaptive-Computation-Machine/dp/0262035618)，該書也可在線獲取：[https://www.deeplearningbook.org/](https://www.deeplearningbook.org/)。
|
||||
|
||||
如果想要輕鬆入門 _雲端 AI_ 主題，可以考慮參加 [Get started with artificial intelligence on Azure](https://docs.microsoft.com/learn/paths/get-started-with-artificial-intelligence-on-azure/?WT.mc_id=academic-77998-bethanycheum) 的學習路徑。
|
||||
如果您想對 _雲端 AI_ 主題有一個簡單的入門，可以考慮參加 [在 Azure 上開始人工智能之旅](https://docs.microsoft.com/learn/paths/get-started-with-artificial-intelligence-on-azure/?WT.mc_id=academic-77998-bethanycheum) 學習路徑。
|
||||
|
||||
# 課程內容
|
||||
|
||||
| | 課程連結 | PyTorch/Keras/TensorFlow | 實驗室 |
|
||||
| :-: | :------------------------------------------------------------------------------------------------------------------------------------------: | :---------------------------------------------------------------------------------------------: | ------------------------------------------------------------------------------ |
|
||||
| 0 | [課程設置](./lessons/0-course-setup/setup.md) | [設置開發環境](./lessons/0-course-setup/how-to-run.md) | |
|
||||
| 0 | [課程設置](./lessons/0-course-setup/setup.md) | [設置您的開發環境](./lessons/0-course-setup/how-to-run.md) | |
|
||||
| I | [**AI 簡介**](./lessons/1-Intro/README.md) | | |
|
||||
| 01 | [AI 的簡介與歷史](./lessons/1-Intro/README.md) | - | - |
|
||||
| II | **符號 AI** |
|
||||
| 02 | [知識表示與專家系統](./lessons/2-Symbolic/README.md) | [專家系統](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/2-Symbolic/Animals.ipynb) / [本體論](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/2-Symbolic/FamilyOntology.ipynb) /[概念圖](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/2-Symbolic/MSConceptGraph.ipynb) | |
|
||||
| 02 | [知識表示與專家系統](./lessons/2-Symbolic/README.md) | [專家系統](./lessons/2-Symbolic/Animals.ipynb) / [本體論](./lessons/2-Symbolic/FamilyOntology.ipynb) /[概念圖](./lessons/2-Symbolic/MSConceptGraph.ipynb) | |
|
||||
| III | [**神經網絡簡介**](./lessons/3-NeuralNetworks/README.md) |||
|
||||
| 03 | [感知器](./lessons/3-NeuralNetworks/03-Perceptron/README.md) | [Notebook](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/3-NeuralNetworks/03-Perceptron/Perceptron.ipynb) | [實驗室](./lessons/3-NeuralNetworks/03-Perceptron/lab/README.md) |
|
||||
| 04 | [多層感知器與創建我們自己的框架](./lessons/3-NeuralNetworks/04-OwnFramework/README.md) | [Notebook](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/3-NeuralNetworks/04-OwnFramework/OwnFramework.ipynb) | [實驗室](./lessons/3-NeuralNetworks/04-OwnFramework/lab/README.md) |
|
||||
| 05 | [框架簡介 (PyTorch/TensorFlow) 與過擬合](./lessons/3-NeuralNetworks/05-Frameworks/README.md) | [PyTorch](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/3-NeuralNetworks/05-Frameworks/IntroPyTorch.ipynb) / [Keras](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/3-NeuralNetworks/05-Frameworks/IntroKeras.ipynb) / [TensorFlow](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/3-NeuralNetworks/05-Frameworks/IntroKerasTF.ipynb) | [實驗室](./lessons/3-NeuralNetworks/05-Frameworks/lab/README.md) |
|
||||
| IV | [**計算機視覺**](./lessons/4-ComputerVision/README.md) | [PyTorch](https://docs.microsoft.com/learn/modules/intro-computer-vision-pytorch/?WT.mc_id=academic-77998-cacaste) / [TensorFlow](https://docs.microsoft.com/learn/modules/intro-computer-vision-TensorFlow/?WT.mc_id=academic-77998-cacaste)| [探索 Microsoft Azure 上的計算機視覺](https://learn.microsoft.com/en-us/collections/7w28iy2xrqzdj0?WT.mc_id=academic-77998-bethanycheum) |
|
||||
| 06 | [計算機視覺簡介. OpenCV](./lessons/4-ComputerVision/06-IntroCV/README.md) | [Notebook](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/4-ComputerVision/06-IntroCV/OpenCV.ipynb) | [實驗室](./lessons/4-ComputerVision/06-IntroCV/lab/README.md) |
|
||||
| 07 | [卷積神經網絡](./lessons/4-ComputerVision/07-ConvNets/README.md) & [CNN 架構](./lessons/4-ComputerVision/07-ConvNets/CNN_Architectures.md) | [PyTorch](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/4-ComputerVision/07-ConvNets/ConvNetsPyTorch.ipynb) /[TensorFlow](https://microsoft.github.io/AI-For-Beginners/lessons/4-ComputerVision/07-ConvNets/ConvNetsTF.ipynb) | [實驗室](./lessons/4-ComputerVision/07-ConvNets/lab/README.md) |
|
||||
| 08 | [預訓練網絡與遷移學習](./lessons/4-ComputerVision/08-TransferLearning/README.md) 和 [訓練技巧](./lessons/4-ComputerVision/08-TransferLearning/TrainingTricks.md) | [PyTorch](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/4-ComputerVision/08-TransferLearning/TransferLearningPyTorch.ipynb) / [TensorFlow](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/3-NeuralNetworks/05-Frameworks/IntroKerasTF.ipynb) | [實驗室](./lessons/4-ComputerVision/08-TransferLearning/lab/README.md) |
|
||||
| 09 | [自編碼器與變分自編碼器 (VAEs)](./lessons/4-ComputerVision/09-Autoencoders/README.md) | [PyTorch](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/4-ComputerVision/09-Autoencoders/AutoEncodersPyTorch.ipynb) / [TensorFlow](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/4-ComputerVision/09-Autoencoders/AutoencodersTF.ipynb) | |
|
||||
| 10 | [生成對抗網絡 (GANs) 與藝術風格遷移](./lessons/4-ComputerVision/10-GANs/README.md) | [PyTorch](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/4-ComputerVision/10-GANs/GANPyTorch.ipynb) / [TensorFlow](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/4-ComputerVision/10-GANs/GANTF.ipynb) | |
|
||||
| 11 | [物件檢測](./lessons/4-ComputerVision/11-ObjectDetection/README.md) | [TensorFlow](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/4-ComputerVision/11-ObjectDetection/ObjectDetection.ipynb) | [實驗室](./lessons/4-ComputerVision/11-ObjectDetection/lab/README.md) |
|
||||
| 12 | [語義分割. U-Net](./lessons/4-ComputerVision/12-Segmentation/README.md) | [PyTorch](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/4-ComputerVision/12-Segmentation/SemanticSegmentationPytorch.ipynb) / [TensorFlow](../../(https:/github.com/microsoft/AI-For-Beginners/blob/main/lessons/4-ComputerVision/12-Segmentation/SemanticSegmentationTF.ipynb)) | |
|
||||
| V | [**自然語言處理**](./lessons/5-NLP/README.md) | [PyTorch](https://docs.microsoft.com/learn/modules/intro-natural-language-processing-pytorch/?WT.mc_id=academic-77998-cacaste) /[TensorFlow](https://docs.microsoft.com/learn/modules/intro-natural-language-processing-TensorFlow/?WT.mc_id=academic-77998-cacaste) | [在 Microsoft Azure 上探索自然語言處理](https://learn.microsoft.com/en-us/collections/7w28iy2xrqzdj0?WT.mc_id=academic-77998-bethanycheum)|
|
||||
| 13 | [文本表示. Bow/TF-IDF](./lessons/5-NLP/13-TextRep/README.md) | [PyTorch](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/5-NLP/13-TextRep/TextRepresentationPyTorch.ipynb) / [TensorFlow](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/5-NLP/13-TextRep/TextRepresentationTF.ipynb) | |
|
||||
| 14 | [語義詞嵌入. Word2Vec 和 GloVe](./lessons/5-NLP/14-Embeddings/README.md) | [PyTorch](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/5-NLP/14-Embeddings/EmbeddingsPyTorch.ipynb) / [TensorFlow](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/5-NLP/14-Embeddings/EmbeddingsTF.ipynb) | |
|
||||
| 15 | [語言建模. 訓練自己的嵌入](./lessons/5-NLP/15-LanguageModeling/README.md) | [PyTorch](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/5-NLP/15-LanguageModeling/CBoW-PyTorch.ipynb) / [TensorFlow](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/5-NLP/15-LanguageModeling/CBoW-TF.ipynb) | [實驗室](./lessons/5-NLP/15-LanguageModeling/lab/README.md) |
|
||||
| 16 | [循環神經網絡 (RNN)](./lessons/5-NLP/16-RNN/README.md) | [PyTorch](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/5-NLP/16-RNN/RNNPyTorch.ipynb) / [TensorFlow](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/5-NLP/16-RNN/RNNTF.ipynb) | |
|
||||
| 17 | [生成型循環網絡](./lessons/5-NLP/17-GenerativeNetworks/README.md) | [PyTorch](https://microsoft.github.io/AI-For-Beginners/lessons/5-NLP/17-GenerativeNetworks/GenerativePyTorch.md) / [TensorFlow](https://microsoft.github.io/AI-For-Beginners/lessons/5-NLP/17-GenerativeNetworks/GenerativeTF.md) | [實驗室](./lessons/5-NLP/17-GenerativeNetworks/lab/README.md) |
|
||||
| 18 | [變壓器. BERT.](./lessons/5-NLP/18-Transformers/READMEtransformers.md) | [PyTorch](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/5-NLP/18-Transformers/TransformersPyTorch.ipynb) /[TensorFlow](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/5-NLP/18-Transformers/TransformersTF.ipynb) | |
|
||||
| 19 | [命名實體識別 (NER)](./lessons/5-NLP/19-NER/README.md) | [TensorFlow](https://microsoft.github.io/AI-For-Beginners/lessons/5-NLP/19-NER/NER-TF.ipynb) | [實驗室](./lessons/5-NLP/19-NER/lab/README.md) |
|
||||
| 20 | [大型語言模型、提示編程與少樣本任務](./lessons/5-NLP/20-LangModels/READMELargeLang.md) | [PyTorch](https://microsoft.github.io/AI-For-Beginners/lessons/5-NLP/20-LangModels/GPT-PyTorch.ipynb) | |
|
||||
| 03 | [感知器](./lessons/3-NeuralNetworks/03-Perceptron/README.md) | [筆記本](./lessons/3-NeuralNetworks/03-Perceptron/Perceptron.ipynb) | [實驗室](./lessons/3-NeuralNetworks/03-Perceptron/lab/README.md) |
|
||||
| 04 | [多層感知器與創建我們自己的框架](./lessons/3-NeuralNetworks/04-OwnFramework/README.md) | [筆記本](./lessons/3-NeuralNetworks/04-OwnFramework/OwnFramework.ipynb) | [實驗室](./lessons/3-NeuralNetworks/04-OwnFramework/lab/README.md) |
|
||||
| 05 | [框架介紹 (PyTorch/TensorFlow) 及過度擬合](./lessons/3-NeuralNetworks/05-Frameworks/README.md) | [PyTorch](./lessons/3-NeuralNetworks/05-Frameworks/IntroPyTorch.ipynb) / [Keras](./lessons/3-NeuralNetworks/05-Frameworks/IntroKeras.ipynb) / [TensorFlow](./lessons/3-NeuralNetworks/05-Frameworks/IntroKerasTF.ipynb) | [實驗室](./lessons/3-NeuralNetworks/05-Frameworks/lab/README.md) |
|
||||
| IV | [**電腦視覺**](./lessons/4-ComputerVision/README.md) | [PyTorch](https://docs.microsoft.com/learn/modules/intro-computer-vision-pytorch/?WT.mc_id=academic-77998-cacaste) / [TensorFlow](https://docs.microsoft.com/learn/modules/intro-computer-vision-TensorFlow/?WT.mc_id=academic-77998-cacaste)| [探索 Microsoft Azure 的電腦視覺](https://learn.microsoft.com/en-us/collections/7w28iy2xrqzdj0?WT.mc_id=academic-77998-bethanycheum) |
|
||||
| 06 | [電腦視覺介紹. OpenCV](./lessons/4-ComputerVision/06-IntroCV/README.md) | [筆記本](./lessons/4-ComputerVision/06-IntroCV/OpenCV.ipynb) | [實驗室](./lessons/4-ComputerVision/06-IntroCV/lab/README.md) |
|
||||
| 07 | [卷積神經網絡](./lessons/4-ComputerVision/07-ConvNets/README.md) & [CNN 架構](./lessons/4-ComputerVision/07-ConvNets/CNN_Architectures.md) | [PyTorch](./lessons/4-ComputerVision/07-ConvNets/ConvNetsPyTorch.ipynb) /[TensorFlow](./lessons/4-ComputerVision/07-ConvNets/ConvNetsTF.ipynb) | [實驗室](./lessons/4-ComputerVision/07-ConvNets/lab/README.md) |
|
||||
| 08 | [預訓練網絡及遷移學習](./lessons/4-ComputerVision/08-TransferLearning/README.md) 和 [訓練技巧](./lessons/4-ComputerVision/08-TransferLearning/TrainingTricks.md) | [PyTorch](./lessons/4-ComputerVision/08-TransferLearning/TransferLearningPyTorch.ipynb) / [TensorFlow](./lessons/3-NeuralNetworks/05-Frameworks/IntroKerasTF.ipynb) | [實驗室](./lessons/4-ComputerVision/08-TransferLearning/lab/README.md) |
|
||||
| 09 | [自編碼器及變分自編碼器 (VAEs)](./lessons/4-ComputerVision/09-Autoencoders/README.md) | [PyTorch](./lessons/4-ComputerVision/09-Autoencoders/AutoEncodersPyTorch.ipynb) / [TensorFlow](./lessons/4-ComputerVision/09-Autoencoders/AutoencodersTF.ipynb) | |
|
||||
| 10 | [生成對抗網絡 (GANs) 及藝術風格轉移](./lessons/4-ComputerVision/10-GANs/README.md) | [PyTorch](./lessons/4-ComputerVision/10-GANs/GANPyTorch.ipynb) / [TensorFlow](./lessons/4-ComputerVision/10-GANs/GANTF.ipynb) | |
|
||||
| 11 | [物件檢測](./lessons/4-ComputerVision/11-ObjectDetection/README.md) | [TensorFlow](./lessons/4-ComputerVision/11-ObjectDetection/ObjectDetection.ipynb) | [實驗室](./lessons/4-ComputerVision/11-ObjectDetection/lab/README.md) |
|
||||
| 12 | [語義分割. U-Net](./lessons/4-ComputerVision/12-Segmentation/README.md) | [PyTorch](./lessons/4-ComputerVision/12-Segmentation/SemanticSegmentationPytorch.ipynb) / [TensorFlow](./lessons/4-ComputerVision/12-Segmentation/SemanticSegmentationTF.ipynb) | |
|
||||
| V | [**自然語言處理**](./lessons/5-NLP/README.md) | [PyTorch](https://docs.microsoft.com/learn/modules/intro-natural-language-processing-pytorch/?WT.mc_id=academic-77998-cacaste) /[TensorFlow](https://docs.microsoft.com/learn/modules/intro-natural-language-processing-TensorFlow/?WT.mc_id=academic-77998-cacaste) | [探索 Microsoft Azure 的自然語言處理](https://learn.microsoft.com/en-us/collections/7w28iy2xrqzdj0?WT.mc_id=academic-77998-bethanycheum)|
|
||||
| 13 | [文本表示. Bow/TF-IDF](./lessons/5-NLP/13-TextRep/README.md) | [PyTorch](./lessons/5-NLP/13-TextRep/TextRepresentationPyTorch.ipynb) / [TensorFlow](./lessons/5-NLP/13-TextRep/TextRepresentationTF.ipynb) | |
|
||||
| 14 | [語義詞嵌入. Word2Vec 和 GloVe](./lessons/5-NLP/14-Embeddings/README.md) | [PyTorch](./lessons/5-NLP/14-Embeddings/EmbeddingsPyTorch.ipynb) / [TensorFlow](./lessons/5-NLP/14-Embeddings/EmbeddingsTF.ipynb) | |
|
||||
| 15 | [語言建模. 訓練自己的嵌入](./lessons/5-NLP/15-LanguageModeling/README.md) | [PyTorch](./lessons/5-NLP/15-LanguageModeling/CBoW-PyTorch.ipynb) / [TensorFlow](./lessons/5-NLP/15-LanguageModeling/CBoW-TF.ipynb) | [實驗室](./lessons/5-NLP/15-LanguageModeling/lab/README.md) |
|
||||
| 16 | [循環神經網絡 (RNN)](./lessons/5-NLP/16-RNN/README.md) | [PyTorch](./lessons/5-NLP/16-RNN/RNNPyTorch.ipynb) / [TensorFlow](./lessons/5-NLP/16-RNN/RNNTF.ipynb) | |
|
||||
| 17 | [生成型循環網絡](./lessons/5-NLP/17-GenerativeNetworks/README.md) | [PyTorch](./lessons/5-NLP/17-GenerativeNetworks/GenerativePyTorch.md) / [TensorFlow](./lessons/5-NLP/17-GenerativeNetworks/GenerativeTF.md) | [實驗室](./lessons/5-NLP/17-GenerativeNetworks/lab/README.md) |
|
||||
| 18 | [變壓器. BERT.](./lessons/5-NLP/18-Transformers/READMEtransformers.md) | [PyTorch](./lessons/5-NLP/18-Transformers/TransformersPyTorch.ipynb) /[TensorFlow](./lessons/5-NLP/18-Transformers/TransformersTF.ipynb) | |
|
||||
| 19 | [命名實體識別 (NER)](./lessons/5-NLP/19-NER/README.md) | [TensorFlow](./lessons/5-NLP/19-NER/NER-TF.ipynb) | [實驗室](./lessons/5-NLP/19-NER/lab/README.md) |
|
||||
| 20 | [大型語言模型, 提示編程及少樣本任務](./lessons/5-NLP/20-LangModels/READMELargeLang.md) | [PyTorch](./lessons/5-NLP/20-LangModels/GPT-PyTorch.ipynb) | |
|
||||
| VI | **其他 AI 技術** || |
|
||||
| 21 | [遺傳算法](./lessons/6-Other/21-GeneticAlgorithms/README.md) | [Notebook](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/6-Other/21-GeneticAlgorithms/Genetic.ipynb) | |
|
||||
| 22 | [深度強化學習](./lessons/6-Other/22-DeepRL/README.md) | [PyTorch](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/6-Other/22-DeepRL/CartPole-RL-PyTorch.ipynb) /[TensorFlow](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/6-Other/22-DeepRL/CartPole-RL-TF.ipynb) | [實驗室](./lessons/6-Other/22-DeepRL/lab/README.md) |
|
||||
| 21 | [遺傳算法](./lessons/6-Other/21-GeneticAlgorithms/README.md) | [筆記本](./lessons/6-Other/21-GeneticAlgorithms/Genetic.ipynb) | |
|
||||
| 22 | [深度強化學習](./lessons/6-Other/22-DeepRL/README.md) | [PyTorch](./lessons/6-Other/22-DeepRL/CartPole-RL-PyTorch.ipynb) /[TensorFlow](./lessons/6-Other/22-DeepRL/CartPole-RL-TF.ipynb) | [實驗室](./lessons/6-Other/22-DeepRL/lab/README.md) |
|
||||
| 23 | [多代理系統](./lessons/6-Other/23-MultiagentSystems/README.md) | | |
|
||||
| VII | **AI 倫理** | | |
|
||||
| 24 | [AI 倫理與負責任的 AI](./lessons/7-Ethics/README.md) | [Microsoft Learn: 負責任的 AI 原則](https://docs.microsoft.com/learn/paths/responsible-ai-business-principles/?WT.mc_id=academic-77998-cacaste) | |
|
||||
| 24 | [AI 倫理及負責任的 AI](./lessons/7-Ethics/README.md) | [Microsoft Learn: 負責任的 AI 原則](https://docs.microsoft.com/learn/paths/responsible-ai-business-principles/?WT.mc_id=academic-77998-cacaste) | |
|
||||
| IX | **附加內容** | | |
|
||||
| 25 | [多模態網絡、CLIP 和 VQGAN](./lessons/X-Extras/X1-MultiModal/README.md) | [Notebook](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/X-Extras/X1-MultiModal/Clip.ipynb) | |
|
||||
| 25 | [多模態網絡, CLIP 和 VQGAN](./lessons/X-Extras/X1-MultiModal/README.md) | [筆記本](./lessons/X-Extras/X1-MultiModal/Clip.ipynb) | |
|
||||
|
||||
## 每節課包含
|
||||
|
||||
* 預讀材料
|
||||
* 可執行的 Jupyter Notebook，通常針對特定框架（**PyTorch** 或 **TensorFlow**）。可執行的 Notebook 也包含大量理論材料，因此要理解主題，您需要至少閱讀一個版本的 Notebook（PyTorch 或 TensorFlow）。
|
||||
* **實驗室**（部分主題提供），讓您有機會將所學材料應用於特定問題。
|
||||
* 可執行的 Jupyter 筆記本，通常針對特定框架（**PyTorch** 或 **TensorFlow**）。可執行的筆記本還包含大量理論材料，因此要理解主題，您需要至少完成一個版本的筆記本（PyTorch 或 TensorFlow）。
|
||||
* **實驗室**（部分主題提供），讓您有機會將所學應用於特定問題。
|
||||
* 部分章節包含指向 [**MS Learn**](https://learn.microsoft.com/en-us/collections/7w28iy2xrqzdj0?WT.mc_id=academic-77998-bethanycheum) 模組的鏈接，涵蓋相關主題。
|
||||
|
||||
## 開始學習
|
||||
|
||||
- 我們已經創建了一個 [設置課程](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/0-course-setup/setup.md)，幫助您設置開發環境。
|
||||
- 對於教育工作者，我們還創建了一個 [課程設置指南](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/0-course-setup/for-teachers.md)！
|
||||
- 如何 [在 VSCode 或 Codepace 中運行代碼](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/0-course-setup/how-to-run.md)
|
||||
- 我們已創建 [設置課程](./lessons/0-course-setup/setup.md) 幫助您設置開發環境。
|
||||
- 對於教育工作者，我們還創建了 [課程設置指南](./lessons/0-course-setup/for-teachers.md)。
|
||||
- 如何 [在 VSCode 或 Codepace 中運行代碼](./lessons/0-course-setup/how-to-run.md)
|
||||
|
||||
按照以下步驟操作：
|
||||
請按照以下步驟操作：
|
||||
|
||||
Fork 此倉庫：點擊此頁面右上角的 "Fork" 按鈕。
|
||||
1. Fork 此倉庫：點擊頁面右上角的 "Fork" 按鈕。
|
||||
2. 克隆倉庫：`git clone https://github.com/microsoft/AI-For-Beginners.git`
|
||||
3. 別忘了給此倉庫加星標 (🌟)，以便日後更容易找到。
|
||||
|
||||
克隆此倉庫：`git clone https://github.com/microsoft/AI-For-Beginners.git`
|
||||
## 與其他學員交流
|
||||
|
||||
別忘了給這個倉庫加星標 (🌟)，以便日後更容易找到。
|
||||
加入我們的 [官方 AI Discord 伺服器](https://aka.ms/genai-discord?WT.mc_id=academic-105485-bethanycheum)，與其他學員交流並獲得支持。
|
||||
|
||||
## 與其他學習者交流
|
||||
|
||||
加入我們的 [官方 AI Discord 伺服器](https://aka.ms/genai-discord?WT.mc_id=academic-105485-bethanycheum)，與其他學習者交流並獲得支持。
|
||||
|
||||
如果您在構建過程中有產品反饋或問題，請訪問我們的 [Azure AI Foundry 開發者論壇](https://aka.ms/foundry/forum)
|
||||
如果您在構建過程中有產品反饋或問題，請訪問我們的 [Azure AI Foundry 開發者論壇](https://aka.ms/foundry/forum)。
|
||||
|
||||
## 測驗
|
||||
> **關於測驗的注意事項**：所有測驗都存放在 Quiz-app 資料夾內，路徑為 etc\quiz-app。這些測驗已經從課程中連結起來，測驗應用程式可以在本地執行，或者部署到 Azure；請按照 `quiz-app` 資料夾中的指引操作。測驗內容正在逐步進行本地化。
|
||||
## 尋求協助
|
||||
> **關於測驗的注意事項**：所有測驗都存放在 etc\quiz-app 的 Quiz-app 資料夾中，或者可以在[線上查看](https://ff-quizzes.netlify.app/)。測驗已在課程中連結，測驗應用程式可以在本地執行或部署到 Azure；請按照 `quiz-app` 資料夾中的指示操作。測驗正在逐步進行本地化。
|
||||
## 幫助需求
|
||||
|
||||
有任何建議或發現拼寫或程式碼錯誤？請提出問題或建立拉取請求。
|
||||
你有任何建議，或者發現拼寫或程式碼錯誤嗎？請提交一個 issue 或建立一個 pull request。
|
||||
|
||||
## 特別感謝
|
||||
## 特別鳴謝
|
||||
|
||||
* **✍️ 主要作者：** [Dmitry Soshnikov](http://soshnikov.com)，博士
|
||||
* **🔥 編輯：** [Jen Looper](https://twitter.com/jenlooper)，博士
|
||||
* **🎨 手繪筆記插畫師：** [Tomomi Imura](https://twitter.com/girlie_mac)
|
||||
* **🎨 插畫設計：** [Tomomi Imura](https://twitter.com/girlie_mac)
|
||||
* **✅ 測驗創作者：** [Lateefah Bello](https://github.com/CinnamonXI)，[MLSA](https://studentambassadors.microsoft.com/)
|
||||
* **🙏 核心貢獻者：** [Evgenii Pishchik](https://github.com/Pe4enIks)
|
||||
|
||||
## 其他課程
|
||||
|
||||
我們的團隊還製作了其他課程！查看以下內容：
|
||||
我們的團隊還製作了其他課程！請查看以下內容：
|
||||
|
||||
- [初學者的生成式 AI](https://aka.ms/genai-beginners)
|
||||
- [初學者的生成式 AI .NET](https://github.com/microsoft/Generative-AI-for-beginners-dotnet)
|
||||
- [初學者生成式 AI](https://aka.ms/genai-beginners)
|
||||
- [初學者生成式 AI .NET](https://github.com/microsoft/Generative-AI-for-beginners-dotnet)
|
||||
- [使用 JavaScript 的生成式 AI](https://github.com/microsoft/generative-ai-with-javascript)
|
||||
- [使用 Java 的生成式 AI](https://github.com/microsoft/Generative-AI-for-beginners-java)
|
||||
- [初學者的 AI](https://aka.ms/ai-beginners)
|
||||
- [初學者的數據科學](https://aka.ms/datascience-beginners)
|
||||
- [初學者的機器學習](https://aka.ms/ml-beginners)
|
||||
- [初學者的網絡安全](https://github.com/microsoft/Security-101)
|
||||
- [初學者的網頁開發](https://aka.ms/webdev-beginners)
|
||||
- [初學者的物聯網](https://aka.ms/iot-beginners)
|
||||
- [初學者的 XR 開發](https://github.com/microsoft/xr-development-for-beginners)
|
||||
- [精通 GitHub Copilot 用於代理性使用](https://github.com/microsoft/Mastering-GitHub-Copilot-for-Paired-Programming)
|
||||
- [精通 GitHub Copilot 用於 C#/.NET 開發者](https://github.com/microsoft/mastering-github-copilot-for-dotnet-csharp-developers)
|
||||
- [初學者 AI](https://aka.ms/ai-beginners)
|
||||
- [初學者數據科學](https://aka.ms/datascience-beginners)
|
||||
- [初學者機器學習](https://aka.ms/ml-beginners)
|
||||
- [初學者網絡安全](https://github.com/microsoft/Security-101)
|
||||
- [初學者網頁開發](https://aka.ms/webdev-beginners)
|
||||
- [初學者物聯網](https://aka.ms/iot-beginners)
|
||||
- [初學者 XR 開發](https://github.com/microsoft/xr-development-for-beginners)
|
||||
- [精通 GitHub Copilot 的智能應用](https://github.com/microsoft/Mastering-GitHub-Copilot-for-Paired-Programming)
|
||||
- [精通 GitHub Copilot 的 C#/.NET 開發者](https://github.com/microsoft/mastering-github-copilot-for-dotnet-csharp-developers)
|
||||
- [選擇你的 Copilot 冒險](https://github.com/microsoft/CopilotAdventures)
|
||||
|
||||
---
|
||||
|
||||
**免責聲明**：
|
||||
本文件使用人工智能翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。應以原文文件作為權威來源。對於關鍵資訊，建議尋求專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或誤釋不承擔責任。
|
||||
本文件已使用人工智能翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。儘管我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。原始文件的母語版本應被視為權威來源。對於重要信息，建議使用專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或錯誤解釋概不負責。
|
||||
|
|
@ -0,0 +1,478 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"collapsed": true
|
||||
},
|
||||
"source": [
|
||||
"# 實現動物專家系統\n",
|
||||
"\n",
|
||||
"來自 [AI for Beginners Curriculum](http://github.com/microsoft/ai-for-beginners) 的範例。\n",
|
||||
"\n",
|
||||
"在這個範例中，我們將實現一個簡單的知識型系統，根據一些外觀特徵來判斷動物的種類。該系統可以用以下的 AND-OR 樹來表示（這只是整棵樹的一部分，我們可以輕鬆地添加更多規則）：\n",
|
||||
"\n",
|
||||
"\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 我們自己的專家系統殼層，具備反向推理功能\n",
|
||||
"\n",
|
||||
"讓我們嘗試定義一種基於生成規則的簡單知識表示語言。我們將使用 Python 類作為關鍵字來定義規則。基本上會有三種類型的類別：\n",
|
||||
"* `Ask` 代表需要向使用者提問的問題。它包含一組可能的答案。\n",
|
||||
"* `If` 代表一條規則，它只是用來存儲規則內容的語法糖。\n",
|
||||
"* `AND`/`OR` 是用來表示樹的 AND/OR 分支的類別。它們僅存儲內部的參數列表。為了簡化程式碼，所有功能都定義在父類別 `Content` 中。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 1,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"class Ask():\n",
|
||||
" def __init__(self,choices=['y','n']):\n",
|
||||
" self.choices = choices\n",
|
||||
" def ask(self):\n",
|
||||
" if max([len(x) for x in self.choices])>1:\n",
|
||||
" for i,x in enumerate(self.choices):\n",
|
||||
" print(\"{0}. {1}\".format(i,x),flush=True)\n",
|
||||
" x = int(input())\n",
|
||||
" return self.choices[x]\n",
|
||||
" else:\n",
|
||||
" print(\"/\".join(self.choices),flush=True)\n",
|
||||
" return input()\n",
|
||||
"\n",
|
||||
"class Content():\n",
|
||||
" def __init__(self,x):\n",
|
||||
" self.x=x\n",
|
||||
" \n",
|
||||
"class If(Content):\n",
|
||||
" pass\n",
|
||||
"\n",
|
||||
"class AND(Content):\n",
|
||||
" pass\n",
|
||||
"\n",
|
||||
"class OR(Content):\n",
|
||||
" pass"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"在我們的系統中，工作記憶將包含**事實**列表，作為**屬性-值對**。知識庫可以定義為一個大的字典，將行動（應插入工作記憶中的新事實）映射到條件，表達為AND-OR表達式。此外，一些事實可以被`詢問`。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 2,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"rules = {\n",
|
||||
" 'default': Ask(['y','n']),\n",
|
||||
" 'color' : Ask(['red-brown','black and white','other']),\n",
|
||||
" 'pattern' : Ask(['dark stripes','dark spots']),\n",
|
||||
" 'mammal': If(OR(['hair','gives milk'])),\n",
|
||||
" 'carnivor': If(OR([AND(['sharp teeth','claws','forward-looking eyes']),'eats meat'])),\n",
|
||||
" 'ungulate': If(['mammal',OR(['has hooves','chews cud'])]),\n",
|
||||
" 'bird': If(OR(['feathers',AND(['flies','lies eggs'])])),\n",
|
||||
" 'animal:monkey' : If(['mammal','carnivor','color:red-brown','pattern:dark spots']),\n",
|
||||
" 'animal:tiger' : If(['mammal','carnivor','color:red-brown','pattern:dark stripes']),\n",
|
||||
" 'animal:giraffe' : If(['ungulate','long neck','long legs','pattern:dark spots']),\n",
|
||||
" 'animal:zebra' : If(['ungulate','pattern:dark stripes']),\n",
|
||||
" 'animal:ostrich' : If(['bird','long nech','color:black and white','cannot fly']),\n",
|
||||
" 'animal:pinguin' : If(['bird','swims','color:black and white','cannot fly']),\n",
|
||||
" 'animal:albatross' : If(['bird','flies well'])\n",
|
||||
"}"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"為了進行反向推理，我們將定義 `Knowledgebase` 類別。它將包含：\n",
|
||||
"* 工作的 `記憶體` - 一個字典，用於映射屬性到其值\n",
|
||||
"* 知識庫的 `規則`，格式如上所述\n",
|
||||
"\n",
|
||||
"兩個主要方法是：\n",
|
||||
"* `get` 用於獲取屬性的值，必要時執行推理。例如，`get('color')` 會獲取顏色槽的值（如果需要，它會詢問並將值存儲在工作記憶中以供後續使用）。如果我們詢問 `get('color:blue')`，它會詢問顏色，然後根據顏色返回 `y`/`n` 的值。\n",
|
||||
"* `eval` 執行實際的推理，即遍歷 AND/OR 樹，評估子目標等。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"class KnowledgeBase():\n",
|
||||
" def __init__(self,rules):\n",
|
||||
" self.rules = rules\n",
|
||||
" self.memory = {}\n",
|
||||
" \n",
|
||||
" def get(self,name):\n",
|
||||
" if ':' in name:\n",
|
||||
" k,v = name.split(':')\n",
|
||||
" vv = self.get(k)\n",
|
||||
" return 'y' if v==vv else 'n'\n",
|
||||
" if name in self.memory.keys():\n",
|
||||
" return self.memory[name]\n",
|
||||
" for fld in self.rules.keys():\n",
|
||||
" if fld==name or fld.startswith(name+\":\"):\n",
|
||||
" # print(\" + proving {}\".format(fld))\n",
|
||||
" value = 'y' if fld==name else fld.split(':')[1]\n",
|
||||
" res = self.eval(self.rules[fld],field=name)\n",
|
||||
" if res!='y' and res!='n' and value=='y':\n",
|
||||
" self.memory[name] = res\n",
|
||||
" return res\n",
|
||||
" if res=='y':\n",
|
||||
" self.memory[name] = value\n",
|
||||
" return value\n",
|
||||
" # field is not found, using default\n",
|
||||
" res = self.eval(self.rules['default'],field=name)\n",
|
||||
" self.memory[name]=res\n",
|
||||
" return res\n",
|
||||
" \n",
|
||||
" def eval(self,expr,field=None):\n",
|
||||
" # print(\" + eval {}\".format(expr))\n",
|
||||
" if isinstance(expr,Ask):\n",
|
||||
" print(field)\n",
|
||||
" return expr.ask()\n",
|
||||
" elif isinstance(expr,If):\n",
|
||||
" return self.eval(expr.x)\n",
|
||||
" elif isinstance(expr,AND) or isinstance(expr,list):\n",
|
||||
" expr = expr.x if isinstance(expr,AND) else expr\n",
|
||||
" for x in expr:\n",
|
||||
" if self.eval(x)=='n':\n",
|
||||
" return 'n'\n",
|
||||
" return 'y'\n",
|
||||
" elif isinstance(expr,OR):\n",
|
||||
" for x in expr.x:\n",
|
||||
" if self.eval(x)=='y':\n",
|
||||
" return 'y'\n",
|
||||
" return 'n'\n",
|
||||
" elif isinstance(expr,str):\n",
|
||||
" return self.get(expr)\n",
|
||||
" else:\n",
|
||||
" print(\"Unknown expr: {}\".format(expr))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在讓我們定義我們的動物知識庫並進行諮詢。請注意，此操作將向您提問。您可以通過輸入 `y`/`n` 來回答是非問題，或者通過指定數字（0..N）來回答有多個選項的問題。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"hair\n",
|
||||
"y/n\n",
|
||||
"sharp teeth\n",
|
||||
"y/n\n",
|
||||
"claws\n",
|
||||
"y/n\n",
|
||||
"forward-looking eyes\n",
|
||||
"y/n\n",
|
||||
"color\n",
|
||||
"0. red-brown\n",
|
||||
"1. black and white\n",
|
||||
"2. other\n",
|
||||
"has hooves\n",
|
||||
"y/n\n",
|
||||
"long neck\n",
|
||||
"y/n\n",
|
||||
"long legs\n",
|
||||
"y/n\n",
|
||||
"pattern\n",
|
||||
"0. dark stripes\n",
|
||||
"1. dark spots\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"'giraffe'"
|
||||
]
|
||||
},
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"kb = KnowledgeBase(rules)\n",
|
||||
"kb.get('animal')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 使用 PyKnow 進行前向推理\n",
|
||||
"\n",
|
||||
"在以下的例子中，我們將嘗試使用其中一個知識表示庫 [PyKnow](https://github.com/buguroo/pyknow/) 來實現前向推理。**PyKnow** 是一個用於在 Python 中建立前向推理系統的庫，其設計理念與經典的舊系統 [CLIPS](http://www.clipsrules.net/index.html) 相似。\n",
|
||||
"\n",
|
||||
"我們其實也可以自己實現前向鏈推理，並不會有太大的困難，但簡單的實現通常效率不高。為了更有效地進行規則匹配，會使用一種特殊的算法 [Rete](https://en.wikipedia.org/wiki/Rete_algorithm)。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Collecting git+https://github.com/buguroo/pyknow/\n",
|
||||
" Cloning https://github.com/buguroo/pyknow/ to /tmp/pip-req-build-3cqeulyl\n",
|
||||
" Running command git clone --filter=blob:none --quiet https://github.com/buguroo/pyknow/ /tmp/pip-req-build-3cqeulyl\n",
|
||||
" Resolved https://github.com/buguroo/pyknow/ to commit 48818336f2e9a126f1964f2d8dc22d37ff800fe8\n",
|
||||
" Preparing metadata (setup.py) ... \u001b[?25ldone\n",
|
||||
"\u001b[?25hCollecting frozendict==1.2\n",
|
||||
" Using cached frozendict-1.2.tar.gz (2.6 kB)\n",
|
||||
" Preparing metadata (setup.py) ... \u001b[?25ldone\n",
|
||||
"\u001b[?25hCollecting schema==0.6.7\n",
|
||||
" Using cached schema-0.6.7-py2.py3-none-any.whl (14 kB)\n",
|
||||
"Building wheels for collected packages: pyknow, frozendict\n",
|
||||
" Building wheel for pyknow (setup.py) ... \u001b[?25ldone\n",
|
||||
"\u001b[?25h Created wheel for pyknow: filename=pyknow-1.7.0-py3-none-any.whl size=34228 sha256=b7de5b09292c4007667c72f69b98d5a1b5f7324ff15f9dd8e077c3d5f7aade42\n",
|
||||
" Stored in directory: /tmp/pip-ephem-wheel-cache-k7jpave7/wheels/81/1a/d3/f6c15dbe1955598a37755215f2a10449e7418500d7bd4b9508\n",
|
||||
" Building wheel for frozendict (setup.py) ... \u001b[?25ldone\n",
|
||||
"\u001b[?25h Created wheel for frozendict: filename=frozendict-1.2-py3-none-any.whl size=3148 sha256=2863d55c240d2409cddf05ccfe600591f8478681549fc97555c47c90dc6bb160\n",
|
||||
" Stored in directory: /home/rg/.cache/pip/wheels/49/ac/f8/cb8120244e710bdb479c86198b03c7b08c3c2d3d2bf448fd6e\n",
|
||||
"Successfully built pyknow frozendict\n",
|
||||
"Installing collected packages: schema, frozendict, pyknow\n",
|
||||
"Successfully installed frozendict-1.2 pyknow-1.7.0 schema-0.6.7\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"import sys\n",
|
||||
"!{sys.executable} -m pip install git+https://github.com/buguroo/pyknow/"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 13,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"from pyknow import *\n",
|
||||
"#import pyknow"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們將把我們的系統定義為一個繼承自 `KnowledgeEngine` 的類別。每條規則由帶有 `@Rule` 註解的單獨函數定義，該註解指定了規則應該何時觸發。在規則內，我們可以使用 `declare` 函數添加新事實，添加這些事實將導致前向推理引擎調用更多規則。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 14,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"class Animals(KnowledgeEngine):\n",
|
||||
" @Rule(OR(\n",
|
||||
" AND(Fact('sharp teeth'),Fact('claws'),Fact('forward looking eyes')),\n",
|
||||
" Fact('eats meat')))\n",
|
||||
" def cornivor(self):\n",
|
||||
" self.declare(Fact('carnivor'))\n",
|
||||
" \n",
|
||||
" @Rule(OR(Fact('hair'),Fact('gives milk')))\n",
|
||||
" def mammal(self):\n",
|
||||
" self.declare(Fact('mammal'))\n",
|
||||
"\n",
|
||||
" @Rule(Fact('mammal'),\n",
|
||||
" OR(Fact('has hooves'),Fact('chews cud')))\n",
|
||||
" def hooves(self):\n",
|
||||
" self.declare('ungulate')\n",
|
||||
" \n",
|
||||
" @Rule(OR(Fact('feathers'),AND(Fact('flies'),Fact('lays eggs'))))\n",
|
||||
" def bird(self):\n",
|
||||
" self.declare('bird')\n",
|
||||
" \n",
|
||||
" @Rule(Fact('mammal'),Fact('carnivor'),\n",
|
||||
" Fact(color='red-brown'),\n",
|
||||
" Fact(pattern='dark spots'))\n",
|
||||
" def monkey(self):\n",
|
||||
" self.declare(Fact(animal='monkey'))\n",
|
||||
"\n",
|
||||
" @Rule(Fact('mammal'),Fact('carnivor'),\n",
|
||||
" Fact(color='red-brown'),\n",
|
||||
" Fact(pattern='dark stripes'))\n",
|
||||
" def tiger(self):\n",
|
||||
" self.declare(Fact(animal='tiger'))\n",
|
||||
"\n",
|
||||
" @Rule(Fact('ungulate'),\n",
|
||||
" Fact('long neck'),\n",
|
||||
" Fact('long legs'),\n",
|
||||
" Fact(pattern='dark spots'))\n",
|
||||
" def giraffe(self):\n",
|
||||
" self.declare(Fact(animal='giraffe'))\n",
|
||||
"\n",
|
||||
" @Rule(Fact('ungulate'),\n",
|
||||
" Fact(pattern='dark stripes'))\n",
|
||||
" def zebra(self):\n",
|
||||
" self.declare(Fact(animal='zebra'))\n",
|
||||
"\n",
|
||||
" @Rule(Fact('bird'),\n",
|
||||
" Fact('long neck'),\n",
|
||||
" Fact('cannot fly'),\n",
|
||||
" Fact(color='black and white'))\n",
|
||||
" def straus(self):\n",
|
||||
" self.declare(Fact(animal='ostrich'))\n",
|
||||
"\n",
|
||||
" @Rule(Fact('bird'),\n",
|
||||
" Fact('swims'),\n",
|
||||
" Fact('cannot fly'),\n",
|
||||
" Fact(color='black and white'))\n",
|
||||
" def pinguin(self):\n",
|
||||
" self.declare(Fact(animal='pinguin'))\n",
|
||||
"\n",
|
||||
" @Rule(Fact('bird'),\n",
|
||||
" Fact('flies well'))\n",
|
||||
" def albatros(self):\n",
|
||||
" self.declare(Fact(animal='albatross'))\n",
|
||||
" \n",
|
||||
" @Rule(Fact(animal=MATCH.a))\n",
|
||||
" def print_result(self,a):\n",
|
||||
" print('Animal is {}'.format(a))\n",
|
||||
" \n",
|
||||
" def factz(self,l):\n",
|
||||
" for x in l:\n",
|
||||
" self.declare(x)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"一旦我們定義了一個知識庫，我們會用一些初始事實填充工作記憶，然後調用 `run()` 方法來執行推理。結果你可以看到新的推導事實被添加到工作記憶中，包括關於動物的最終事實（如果我們正確設置了所有初始事實）。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 15,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Animal is tiger\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"FactList([(0, InitialFact()),\n",
|
||||
" (1, Fact(color='red-brown')),\n",
|
||||
" (2, Fact(pattern='dark stripes')),\n",
|
||||
" (3, Fact('sharp teeth')),\n",
|
||||
" (4, Fact('claws')),\n",
|
||||
" (5, Fact('forward looking eyes')),\n",
|
||||
" (6, Fact('gives milk')),\n",
|
||||
" (7, Fact('mammal')),\n",
|
||||
" (8, Fact('carnivor')),\n",
|
||||
" (9, Fact(animal='tiger'))])"
|
||||
]
|
||||
},
|
||||
"execution_count": 15,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"ex1 = Animals()\n",
|
||||
"ex1.reset()\n",
|
||||
"ex1.factz([\n",
|
||||
" Fact(color='red-brown'),\n",
|
||||
" Fact(pattern='dark stripes'),\n",
|
||||
" Fact('sharp teeth'),\n",
|
||||
" Fact('claws'),\n",
|
||||
" Fact('forward looking eyes'),\n",
|
||||
" Fact('gives milk')])\n",
|
||||
"ex1.run()\n",
|
||||
"ex1.facts"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": []
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件已使用人工智能翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。應以原文文件作為權威來源。對於關鍵資訊，建議尋求專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或錯誤解釋概不負責。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"kernelspec": {
|
||||
"display_name": "Python 3.7.4 64-bit (conda)",
|
||||
"metadata": {
|
||||
"interpreter": {
|
||||
"hash": "86193a1ab0ba47eac1c69c1756090baa3b420b3eea7d4aafab8b85f8b312f0c5"
|
||||
}
|
||||
},
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.11.2"
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "ab2bd97b0453415b89a469284609a8ce",
|
||||
"translation_date": "2025-08-31T10:08:03+00:00",
|
||||
"source_file": "lessons/2-Symbolic/Animals.ipynb",
|
||||
"language_code": "hk"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 2
|
||||
}
|
||||
|
|
@ -0,0 +1,595 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"collapsed": true
|
||||
},
|
||||
"source": [
|
||||
"# 家庭關係本體\n",
|
||||
"\n",
|
||||
"這個例子是 [AI for Beginners Curriculum](http://github.com/microsoft/ai-for-beginners) 的一部分，靈感來自於[這篇博客文章](https://habr.com/post/270857/)。\n",
|
||||
"\n",
|
||||
"我總是覺得記住家庭成員之間的不同關係很困難。在這個例子中，我們將使用一個定義家庭關係的本體，以及實際的家譜樹，展示如何通過自動推理來找出所有親屬。\n",
|
||||
"\n",
|
||||
"### 獲取家譜樹\n",
|
||||
"\n",
|
||||
"作為例子，我們將使用[羅曼諾夫沙皇家族](https://en.wikipedia.org/wiki/House_of_Romanov)的家譜樹。描述家庭關係最常見的格式是 [GEDCOM](https://en.wikipedia.org/wiki/GEDCOM)。我們將使用 GEDCOM 格式的羅曼諾夫家族樹：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 1,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"0 HEAD\n",
|
||||
"1 CHAR UTF8\n",
|
||||
"1 GEDC\n",
|
||||
"2 VERS 5.5\n",
|
||||
"0 @0@ INDI\n",
|
||||
"1 NAME Mihail Fedorovich /Romanov/\n",
|
||||
"1 SEX M\n",
|
||||
"1 BIRT\n",
|
||||
"2 DATE 1613\n",
|
||||
"1 DEAT \n",
|
||||
"2 DATE 1645\n",
|
||||
"1 FAMS @41@\n",
|
||||
"0 @1@ INDI\n",
|
||||
"1 NAME Evdokija Lukjanovna /Streshneva/\n",
|
||||
"1 SEX F\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"!head -15 data/tsars.ged"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"要使用 GEDCOM 文件，我們可以使用 `python-gedcom` 庫：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 2,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Collecting python-gedcom\n",
|
||||
" Downloading python_gedcom-1.0.0-py2.py3-none-any.whl (35 kB)\n",
|
||||
"Installing collected packages: python-gedcom\n",
|
||||
"Successfully installed python-gedcom-1.0.0\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"import sys\n",
|
||||
"!{sys.executable} -m pip install python-gedcom"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"這個庫消除了文件解析的一些技術問題，但它仍然讓我們能夠以相當低層次的方式訪問樹中的所有個人和家庭。以下是我們如何解析文件並顯示所有個人列表的方法：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"from gedcom.parser import Parser\n",
|
||||
"from gedcom.element.individual import IndividualElement\n",
|
||||
"from gedcom.element.family import FamilyElement\n",
|
||||
"g = Parser()\n",
|
||||
"g.parse_file('data/tsars.ged')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {
|
||||
"scrolled": true,
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"[('@0@', ('Mihail Fedorovich', 'Romanov')),\n",
|
||||
" ('@1@', ('Evdokija Lukjanovna', 'Streshneva')),\n",
|
||||
" ('@2@', ('Aleksej Mihajlovich', 'Romanov')),\n",
|
||||
" ('@3@', ('Marija Ilinichna', 'Miloslavskaja')),\n",
|
||||
" ('@4@', ('Natalja Kirillovna', 'Naryshkina')),\n",
|
||||
" ('@5@', ('Marfa Matveevna', 'Apraksina')),\n",
|
||||
" ('@6@', ('Fedor Alekseevich', 'Romanov')),\n",
|
||||
" ('@7@', ('Sofja Aleksevna', 'Romanova')),\n",
|
||||
" ('@8@', ('Ivan V Alekseevich', 'Romanov')),\n",
|
||||
" ('@9@', ('Praskovja Fedorovna', 'Saltykova')),\n",
|
||||
" ('@10@', ('Ekaterina Ivanovna', 'Romanova')),\n",
|
||||
" ('@11@', ('Anna Ivanovna', 'Romanova')),\n",
|
||||
" ('@12@', ('Fridrih Vilgelm', 'Kurlandskij')),\n",
|
||||
" ('@13@', ('Karl Leopold', 'Meklenburg-Shverinskij')),\n",
|
||||
" ('@14@', ('Anna Leopoldovna', 'Meklenburg-Shverinskaja')),\n",
|
||||
" ('@15@', ('Anton Ulrih', 'Braunshvejg-Volfenbjuttelskij')),\n",
|
||||
" ('@16@', ('Ivan VI Antonovich', 'Braunshvejg-Volfenbjuttelskij')),\n",
|
||||
" ('@17@', ('Petr I Alekseevich', 'Romanov')),\n",
|
||||
" ('@18@', ('Evdokija Fedorovna', 'Lopuhina')),\n",
|
||||
" ('@19@', ('Ekaterina I Alekseevna', 'Mihajlova')),\n",
|
||||
" ('@20@', ('Aleksej Petrovich', 'Romanov')),\n",
|
||||
" ('@21@', ('Sharlotta Kristina', 'Braunshvejg-Volfenbjuttelskaja')),\n",
|
||||
" ('@22@', ('Petr II Alekseevich', 'Romanov')),\n",
|
||||
" ('@23@', ('Anna Petrovna', 'Romanova')),\n",
|
||||
" ('@24@', ('Elizaveta Petrovna', 'Romanova')),\n",
|
||||
" ('@25@', ('Karl Fridrih', 'Golshtejn-Gottorpskij')),\n",
|
||||
" ('@26@', ('Petr III Fedorovich', 'Romanov')),\n",
|
||||
" ('@27@', ('Ekaterina II', 'Alekseevna')),\n",
|
||||
" ('@28@', ('Pavel I Petrovich', 'Romanov')),\n",
|
||||
" ('@29@', ('Natalja Alekseevna', 'Gessen-Darmshtadskaja')),\n",
|
||||
" ('@30@', ('Marija Fedorovna', 'Vjurtembergskaja')),\n",
|
||||
" ('@31@', ('Aleksandr I Pavlovich', 'Romanov')),\n",
|
||||
" ('@32@', ('Elizaveta Alekseevna', 'Baden-Durlahskaja')),\n",
|
||||
" ('@33@', ('Nikolaj I Pavlovich', 'Romanov')),\n",
|
||||
" ('@34@', ('Aleksandra Fedorovna', 'Prusskaja')),\n",
|
||||
" ('@35@', ('Aleksandr II Nikolaevich', 'Romanov')),\n",
|
||||
" ('@36@', ('Marija Aleksandrovna', 'Gessenskaja')),\n",
|
||||
" ('@37@', ('Aleksandr III Aleksandrovich', 'Romanov')),\n",
|
||||
" ('@38@', ('Marija Fedorovna', 'Datskaja')),\n",
|
||||
" ('@39@', ('Nikolaj II Aleksandrovich', 'Romanov')),\n",
|
||||
" ('@40@', ('Aleksandra Fedorovna', 'Gessenskaja'))]"
|
||||
]
|
||||
},
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"d = g.get_element_dictionary()\n",
|
||||
"[ (k,v.get_name()) for k,v in d.items() if isinstance(v,IndividualElement)]"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"以下是我們如何獲取有關家庭的信息。請注意，這會給我們一個**標識符**的列表，如果我們想要更清楚的資訊，需要將它們轉換為名字：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"[('@41@', ['@0@', '@1@', '@2@']),\n",
|
||||
" ('@42@', ['@2@', '@3@', '@6@', '@7@', '@8@']),\n",
|
||||
" ('@43@', ['@8@', '@9@', '@10@', '@11@']),\n",
|
||||
" ('@44@', ['@13@', '@10@', '@14@']),\n",
|
||||
" ('@45@', ['@15@', '@14@', '@16@']),\n",
|
||||
" ('@46@', ['@2@', '@4@', '@17@']),\n",
|
||||
" ('@47@', ['@17@', '@18@', '@20@']),\n",
|
||||
" ('@48@', ['@20@', '@21@', '@22@']),\n",
|
||||
" ('@49@', ['@17@', '@19@', '@23@', '@24@']),\n",
|
||||
" ('@50@', ['@25@', '@23@', '@26@']),\n",
|
||||
" ('@51@', ['@26@', '@27@', '@28@']),\n",
|
||||
" ('@52@', ['@28@', '@30@', '@31@', '@33@']),\n",
|
||||
" ('@53@', ['@33@', '@34@', '@35@']),\n",
|
||||
" ('@54@', ['@35@', '@36@', '@37@']),\n",
|
||||
" ('@55@', ['@37@', '@38@', '@39@'])]"
|
||||
]
|
||||
},
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"d = g.get_element_dictionary()\n",
|
||||
"[ (k,[x.get_value() for x in v.get_child_elements()]) for k,v in d.items() if isinstance(v,FamilyElement)]"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"### 獲取家庭本體\n",
|
||||
"\n",
|
||||
"接下來，我們來看看[家庭本體](https://raw.githubusercontent.com/blokhin/genealogical-trees/master/data/header.ttl)，它被定義為一組語義網三元組。這個本體定義了例如 `isUncleOf`、`isCousinOf` 等許多關係。所有這些關係都是基於基本謂詞 `isMotherOf`、`isFatherOf`、`isBrotherOf` 和 `isSisterOf` 來定義的。我們將使用自動推理來根據本體推導出所有其他關係。\n",
|
||||
"\n",
|
||||
"以下是 `isAuntOf` 屬性的示例定義，它被定義為 `isSisterOf` 和 `isParentOf` 的組合（*姨母是某人父母的姐妹*）。\n",
|
||||
"\n",
|
||||
"```\n",
|
||||
"fhkb:isAuntOf a owl:ObjectProperty ;\n",
|
||||
" rdfs:domain fhkb:Woman ;\n",
|
||||
" rdfs:range fhkb:Person ;\n",
|
||||
" owl:propertyChainAxiom ( fhkb:isSisterOf fhkb:isParentOf ) .\n",
|
||||
"```\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 6,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"@prefix fhkb: <http://www.example.com/genealogy.owl#> .\n",
|
||||
"@prefix owl: <http://www.w3.org/2002/07/owl#> .\n",
|
||||
"@prefix rdf: <http://www.w3.org/1999/02/22-rdf-syntax-ns#> .\n",
|
||||
"@prefix rdfs: <http://www.w3.org/2000/01/rdf-schema#> .\n",
|
||||
"@prefix xml: <http://www.w3.org/XML/1998/namespace> .\n",
|
||||
"@prefix xsd: <http://www.w3.org/2001/XMLSchema#> .\n",
|
||||
"\n",
|
||||
"<http://www.example.com/genealogy.owl#> a owl:Ontology .\n",
|
||||
"\n",
|
||||
"fhkb:DomainEntity a owl:Class .\n",
|
||||
"\n",
|
||||
"fhkb:Man a owl:Class ;\n",
|
||||
" owl:equivalentClass [ a owl:Class ;\n",
|
||||
" owl:intersectionOf ( fhkb:Person [ a owl:Restriction ;\n",
|
||||
" owl:onProperty fhkb:hasSex ;\n",
|
||||
" owl:someValuesFrom fhkb:Male ] ) ] .\n",
|
||||
"\n",
|
||||
"fhkb:Woman a owl:Class ;\n",
|
||||
" owl:equivalentClass [ a owl:Class ;\n",
|
||||
" owl:intersectionOf ( fhkb:Person [ a owl:Restriction ;\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"!head -20 data/onto.ttl"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"### 建構推理本體\n",
|
||||
"\n",
|
||||
"為了簡化，我們將建立一個本體文件，該文件將包含來自家庭本體的原始規則，以及來自我們 GEDCOM 文件的個人事實。我們會逐一檢視 GEDCOM 文件中的家庭和個人資訊，並將其轉換為三元組。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 7,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"!cp data/onto.ttl .\n",
|
||||
"\n",
|
||||
"gedcom_dict = g.get_element_dictionary()\n",
|
||||
"individuals, marriages = {}, {}\n",
|
||||
"\n",
|
||||
"def term2id(el):\n",
|
||||
" return \"i\" + el.get_pointer().replace('@', '').lower()\n",
|
||||
"\n",
|
||||
"out = open(\"onto.ttl\",\"a\")\n",
|
||||
"\n",
|
||||
"for k, v in gedcom_dict.items():\n",
|
||||
" if isinstance(v,IndividualElement):\n",
|
||||
" children, siblings = set(), set()\n",
|
||||
" idx = term2id(v)\n",
|
||||
"\n",
|
||||
" title = v.get_name()[0] + \" \" + v.get_name()[1]\n",
|
||||
" title = title.replace('\"', '').replace('[', '').replace(']', '').replace('(', '').replace(')', '').strip()\n",
|
||||
"\n",
|
||||
" own_families = g.get_families(v, 'FAMS')\n",
|
||||
" for fam in own_families:\n",
|
||||
" children |= set(term2id(i) for i in g.get_family_members(fam, \"CHIL\"))\n",
|
||||
"\n",
|
||||
" parent_families = g.get_families(v, 'FAMC')\n",
|
||||
" if len(parent_families):\n",
|
||||
" for member in g.get_family_members(parent_families[0], \"CHIL\"): # NB adoptive families i.e len(parent_families)>1 are not considered (TODO?)\n",
|
||||
" if member.get_pointer() == v.get_pointer():\n",
|
||||
" continue\n",
|
||||
" siblings.add(term2id(member))\n",
|
||||
"\n",
|
||||
" if idx in individuals:\n",
|
||||
" children |= individuals[idx].get('children', set())\n",
|
||||
" siblings |= individuals[idx].get('siblings', set())\n",
|
||||
" individuals[idx] = {'sex': v.get_gender().lower(), 'children': children, 'siblings': siblings, 'title': title}\n",
|
||||
"\n",
|
||||
" elif isinstance(v,FamilyElement):\n",
|
||||
" wife, husb, children = None, None, set()\n",
|
||||
" children = set(term2id(i) for i in g.get_family_members(v, \"CHIL\"))\n",
|
||||
"\n",
|
||||
" try:\n",
|
||||
" wife = g.get_family_members(v, \"WIFE\")[0]\n",
|
||||
" wife = term2id(wife)\n",
|
||||
" if wife in individuals: individuals[wife]['children'] |= children\n",
|
||||
" else: individuals[wife] = {'children': children}\n",
|
||||
" except IndexError: pass\n",
|
||||
" try:\n",
|
||||
" husb = g.get_family_members(v, \"HUSB\")[0]\n",
|
||||
" husb = term2id(husb)\n",
|
||||
" if husb in individuals: individuals[husb]['children'] |= children\n",
|
||||
" else: individuals[husb] = {'children': children}\n",
|
||||
" except IndexError: pass\n",
|
||||
"\n",
|
||||
" if wife and husb: marriages[wife + husb] = (term2id(v), wife, husb)\n",
|
||||
"\n",
|
||||
"for idx, val in individuals.items():\n",
|
||||
" added_terms = ''\n",
|
||||
" if val['sex'] == 'f':\n",
|
||||
" parent_predicate, sibl_predicate = \"isMotherOf\", \"isSisterOf\"\n",
|
||||
" else:\n",
|
||||
" parent_predicate, sibl_predicate = \"isFatherOf\", \"isBrotherOf\"\n",
|
||||
" if len(val['children']):\n",
|
||||
" added_terms += \" ;\\n fhkb:\" + parent_predicate + \" \" + \", \".join([\"fhkb:\" + i for i in val['children']])\n",
|
||||
" if len(val['siblings']):\n",
|
||||
" added_terms += \" ;\\n fhkb:\" + sibl_predicate + \" \" + \", \".join([\"fhkb:\" + i for i in val['siblings']])\n",
|
||||
" out.write(\"fhkb:%s a owl:NamedIndividual, owl:Thing%s ;\\n rdfs:label \\\"%s\\\" .\\n\" % (idx, added_terms, val['title']))\n",
|
||||
"\n",
|
||||
"for k, v in marriages.items():\n",
|
||||
" out.write(\"fhkb:%s a owl:NamedIndividual, owl:Thing ;\\n fhkb:hasFemalePartner fhkb:%s ;\\n fhkb:hasMalePartner fhkb:%s .\\n\" % v)\n",
|
||||
"\n",
|
||||
"out.write(\"[] a owl:AllDifferent ;\\n owl:distinctMembers (\")\n",
|
||||
"for idx in individuals.keys():\n",
|
||||
" out.write(\" fhkb:\" + idx)\n",
|
||||
"for k, v in marriages.items():\n",
|
||||
" out.write(\" fhkb:\" + v[0])\n",
|
||||
"out.write(\" ) .\")\n",
|
||||
"out.close()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 8,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
" fhkb:hasFemalePartner fhkb:i34 ;\n",
|
||||
" fhkb:hasMalePartner fhkb:i33 .\n",
|
||||
"fhkb:i54 a owl:NamedIndividual, owl:Thing ;\n",
|
||||
" fhkb:hasFemalePartner fhkb:i36 ;\n",
|
||||
" fhkb:hasMalePartner fhkb:i35 .\n",
|
||||
"fhkb:i55 a owl:NamedIndividual, owl:Thing ;\n",
|
||||
" fhkb:hasFemalePartner fhkb:i38 ;\n",
|
||||
" fhkb:hasMalePartner fhkb:i37 .\n",
|
||||
"[] a owl:AllDifferent ;\n",
|
||||
" owl:distinctMembers ( fhkb:i0 fhkb:i1 fhkb:i2 fhkb:i3 fhkb:i4 fhkb:i5 fhkb:i6 fhkb:i7 fhkb:i8 fhkb:i9 fhkb:i10 fhkb:i11 fhkb:i12 fhkb:i13 fhkb:i14 fhkb:i15 fhkb:i16 fhkb:i17 fhkb:i18 fhkb:i19 fhkb:i20 fhkb:i21 fhkb:i22 fhkb:i23 fhkb:i24 fhkb:i25 fhkb:i26 fhkb:i27 fhkb:i28 fhkb:i29 fhkb:i30 fhkb:i31 fhkb:i32 fhkb:i33 fhkb:i34 fhkb:i35 fhkb:i36 fhkb:i37 fhkb:i38 fhkb:i39 fhkb:i40 fhkb:i41 fhkb:i42 fhkb:i43 fhkb:i44 fhkb:i45 fhkb:i46 fhkb:i47 fhkb:i48 fhkb:i49 fhkb:i50 fhkb:i51 fhkb:i52 fhkb:i53 fhkb:i54 fhkb:i55 ) ."
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"!tail onto.ttl"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"### 執行推理\n",
|
||||
"\n",
|
||||
"現在我們希望能夠使用這個本體進行推理和查詢。我們將使用 [RDFLib](https://github.com/RDFLib)，這是一個用於以不同格式讀取 RDF 圖形、查詢等的庫。\n",
|
||||
"\n",
|
||||
"至於邏輯推理，我們將使用 [OWL-RL](https://github.com/RDFLib/OWL-RL) 庫，這個庫允許我們建立 RDF 圖形的**閉包**，即加入所有可以推導出的概念和關係。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 10,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Requirement already satisfied: rdflib in /home/rg/anaconda3/envs/ai4beg/lib/python3.11/site-packages (6.3.2)\n",
|
||||
"Requirement already satisfied: isodate<0.7.0,>=0.6.0 in /home/rg/anaconda3/envs/ai4beg/lib/python3.11/site-packages (from rdflib) (0.6.1)\n",
|
||||
"Requirement already satisfied: pyparsing<4,>=2.1.0 in /home/rg/anaconda3/envs/ai4beg/lib/python3.11/site-packages (from rdflib) (3.0.9)\n",
|
||||
"Requirement already satisfied: six in /home/rg/anaconda3/envs/ai4beg/lib/python3.11/site-packages (from isodate<0.7.0,>=0.6.0->rdflib) (1.16.0)\n",
|
||||
"Collecting git+https://github.com/RDFLib/OWL-RL.git\n",
|
||||
" Cloning https://github.com/RDFLib/OWL-RL.git to /tmp/pip-req-build-lbfzwi3m\n",
|
||||
" Running command git clone --filter=blob:none --quiet https://github.com/RDFLib/OWL-RL.git /tmp/pip-req-build-lbfzwi3m\n",
|
||||
" Resolved https://github.com/RDFLib/OWL-RL.git to commit a77e1791b88b54aace609bc6000aac14c7add4ff\n",
|
||||
" Preparing metadata (setup.py) ... \u001b[?25ldone\n",
|
||||
"\u001b[?25hRequirement already satisfied: rdflib>=6.0.2 in /home/rg/anaconda3/envs/ai4beg/lib/python3.11/site-packages (from owlrl==6.0.2) (6.3.2)\n",
|
||||
"Requirement already satisfied: isodate<0.7.0,>=0.6.0 in /home/rg/anaconda3/envs/ai4beg/lib/python3.11/site-packages (from rdflib>=6.0.2->owlrl==6.0.2) (0.6.1)\n",
|
||||
"Requirement already satisfied: pyparsing<4,>=2.1.0 in /home/rg/anaconda3/envs/ai4beg/lib/python3.11/site-packages (from rdflib>=6.0.2->owlrl==6.0.2) (3.0.9)\n",
|
||||
"Requirement already satisfied: six in /home/rg/anaconda3/envs/ai4beg/lib/python3.11/site-packages (from isodate<0.7.0,>=0.6.0->rdflib>=6.0.2->owlrl==6.0.2) (1.16.0)\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"!{sys.executable} -m pip install rdflib\n",
|
||||
"!{sys.executable} -m pip install git+https://github.com/RDFLib/OWL-RL.git"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"讓我們打開本體文件，看看它包含多少個三元組：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 11,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Triplets found:669\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"import rdflib\n",
|
||||
"from owlrl import DeductiveClosure, OWLRL_Extension\n",
|
||||
"\n",
|
||||
"g = rdflib.Graph()\n",
|
||||
"g.parse(\"onto.ttl\", format=\"turtle\")\n",
|
||||
"\n",
|
||||
"print(\"Triplets found:%d\" % len(g))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在讓我們建立閉包，看看三元組的數量如何增加：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 12,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Triplets after inference:4246\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"DeductiveClosure(OWLRL_Extension).expand(g)\n",
|
||||
"print(\"Triplets after inference:%d\" % len(g))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"### 查詢親屬關係\n",
|
||||
"\n",
|
||||
"現在我們可以查詢圖譜，來查看人與人之間的不同關係。我們可以結合使用 **SPARQL** 語言和 `query` 方法。在我們的例子中，讓我們看看家譜中所有的 **叔叔**：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 13,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Fedor Alekseevich Romanov is uncle of Ekaterina Ivanovna Romanova\n",
|
||||
"Aleksandr I Pavlovich Romanov is uncle of Aleksandr II Nikolaevich Romanov\n",
|
||||
"Fedor Alekseevich Romanov is uncle of Anna Ivanovna Romanova\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"qres = g.query(\n",
|
||||
" \"\"\"SELECT DISTINCT ?aname ?bname\n",
|
||||
" WHERE {\n",
|
||||
" ?a fhkb:isUncleOf ?b .\n",
|
||||
" ?a rdfs:label ?aname .\n",
|
||||
" ?b rdfs:label ?bname .\n",
|
||||
" }\"\"\")\n",
|
||||
"\n",
|
||||
"for row in qres:\n",
|
||||
" print(\"%s is uncle of %s\" % row)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"隨意嘗試不同的家庭關係。例如，你可以查看 `isAncestorOf` 關係，它會遞歸地定義某個人的所有祖先。\n",
|
||||
"\n",
|
||||
"最後，讓我們清理一下吧！\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 14,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"!rm onto.ttl"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": []
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件已使用人工智能翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。應以原文文件作為權威來源。對於關鍵資訊，建議尋求專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或誤釋不承擔責任。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"interpreter": {
|
||||
"hash": "86193a1ab0ba47eac1c69c1756090baa3b420b3eea7d4aafab8b85f8b312f0c5"
|
||||
},
|
||||
"kernelspec": {
|
||||
"display_name": "Python 3.6",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.11.2"
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "6537d5597320e27b6052b4377b8ff8bb",
|
||||
"translation_date": "2025-08-31T10:06:07+00:00",
|
||||
"source_file": "lessons/2-Symbolic/FamilyOntology.ipynb",
|
||||
"language_code": "hk"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 2
|
||||
}
|
||||
|
|
@ -0,0 +1,548 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"collapsed": true
|
||||
},
|
||||
"source": [
|
||||
"## Microsoft 概念圖\n",
|
||||
"\n",
|
||||
"[Microsoft 概念圖](https://concept.research.microsoft.com/) 是一個從互聯網挖掘而來的大型術語分類，包含概念之間的「是-屬於」關係。\n",
|
||||
"\n",
|
||||
"概念圖有兩種形式可供使用：\n",
|
||||
" * 可下載的大型文本文件\n",
|
||||
" * REST API\n",
|
||||
"\n",
|
||||
"統計數據：\n",
|
||||
" * 5401933 個獨特概念\n",
|
||||
" * 12551613 個獨特實例\n",
|
||||
" * 87603947 條「是-屬於」關係\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 使用網絡服務\n",
|
||||
"\n",
|
||||
"網絡服務提供不同的調用，用於估算某個概念屬於不同群組的概率。更多資訊可參考[這裡](https://concept.research.microsoft.com/Home/Api)。 \n",
|
||||
"以下是範例 URL：`https://concept.research.microsoft.com/api/Concept/ScoreByProb?instance=microsoft&topK=10`\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 7,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"{'company': 0.6105356614382954,\n",
|
||||
" 'vendor': 0.08858636677518003,\n",
|
||||
" 'client': 0.048239124001183784,\n",
|
||||
" 'firm': 0.045476965571668145,\n",
|
||||
" 'large company': 0.043109401203511886,\n",
|
||||
" 'organization': 0.043010752688172046,\n",
|
||||
" 'corporation': 0.035908059583703265,\n",
|
||||
" 'brand': 0.03383644076156654,\n",
|
||||
" 'software company': 0.027522935779816515,\n",
|
||||
" 'technology company': 0.023774292196902438}"
|
||||
]
|
||||
},
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"import urllib\n",
|
||||
"import json\n",
|
||||
"import ssl\n",
|
||||
"\n",
|
||||
"def http(x):\n",
|
||||
" ssl._create_default_https_context = ssl._create_unverified_context\n",
|
||||
" response = urllib.request.urlopen(x)\n",
|
||||
" data = response.read()\n",
|
||||
" return data.decode('utf-8')\n",
|
||||
"\n",
|
||||
"def query(x):\n",
|
||||
" return json.loads(http(\"https://concept.research.microsoft.com/api/Concept/ScoreByProb?instance={}&topK=10\".format(urllib.parse.quote(x))))\n",
|
||||
"\n",
|
||||
"query('microsoft')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"讓我們嘗試使用父概念對新聞標題進行分類。要獲取新聞標題，我們將使用 [NewsApi.org](http://newsapi.org) 服務。您需要獲取自己的 API 密鑰才能使用該服務——請訪問網站並註冊免費開發者計劃。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 20,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"newsapi_key = '<your API key here>'\n",
|
||||
"def get_news(country='us'):\n",
|
||||
" res = json.loads(http(\"https://newsapi.org/v2/top-headlines?country={0}&apiKey={1}\".format(country,newsapi_key)))\n",
|
||||
" return res['articles']\n",
|
||||
"\n",
|
||||
"all_titles = [x['title'] for x in get_news('us')+get_news('gb')]"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 21,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"['Covid-19 Live Updates: Vaccines and Boosters News - The New York Times',\n",
|
||||
" 'Ukrainians Flee Mariupol as Russian Forces Push to Take Port City - The Wall Street Journal',\n",
|
||||
" 'Bond Yields Jump, Stock Futures Rise After Powell Says Fed Is Ready to Be More Aggressive - The Wall Street Journal',\n",
|
||||
" 'Putin critic Alexei Navalny found guilty by Russian court - New York Post ',\n",
|
||||
" \"Supreme Court nominee Ketanji Brown Jackson will face questions at confirmation hearing's second day - CNN\",\n",
|
||||
" '2 teachers killed at Swedish high school, student arrested - ABC News',\n",
|
||||
" 'Clues to Covid-19’s Next Moves Come From Sewers - The Wall Street Journal',\n",
|
||||
" 'Republicans to roll dice by grilling Jackson over child-pornography sentencing decisions | TheHill - The Hill',\n",
|
||||
" '‘Clear sign’ Putin considering using chemical weapons in Ukraine, claims President Biden - The Independent',\n",
|
||||
" 'NASA confirms there are 5,000 planets outside our solar system - Daily Mail',\n",
|
||||
" \"US stocks whipsawed overnight after Fed Chair Powell's remarks - Fox Business\",\n",
|
||||
" \"'We've learned absolutely nothing': Tests could again be in short supply if Covid surges - POLITICO\",\n",
|
||||
" \"Duchess of Cambridge swaps khaki jungle gear for Vampire's Wife dress on Belize trip - Daily Mail\",\n",
|
||||
" 'China searches for victims, flight recorders after first plane crash in 12 years - Reuters',\n",
|
||||
" 'Second superyacht linked to Russian oligarch Abramovich docks in Turkey - Reuters',\n",
|
||||
" 'Live updates: Russia stops talks with Japan over sanctions - The Associated Press - en Español',\n",
|
||||
" 'Powers Remain and Threats Lurk as Women’s Sweet 16 Is Set - The New York Times',\n",
|
||||
" 'Webb Space Telescope Begins Multi-Instrument Alignment - SciTechDaily',\n",
|
||||
" \"UConn vs UCF - NCAA women's tournament second-round highlights - March Madness\",\n",
|
||||
" 'Bucking Republican Trend, Indiana Governor Vetoes Transgender Sports Bill - The New York Times',\n",
|
||||
" \"Maggie Fox dead: Coronation Street and Shameless actress dies after 'sudden accident' - Mirror Online - The Mirror\",\n",
|
||||
" 'China plane crash – live: Search for survivors continues as witness describes moment flight fell from sky - The Independent',\n",
|
||||
" 'Daniel Morgan murder: damning report condemns Met police - The Guardian',\n",
|
||||
" 'What to expect from Rishi Sunak’s Spring Statement - BBC.com',\n",
|
||||
" 'UK and Republic of Ireland in line to host Euro 2028 after no one else bids - The Guardian',\n",
|
||||
" \"Friends beg Vladimir Putin's 'lover' to persuade him to end Ukraine invasion - The Mirror\",\n",
|
||||
" 'Brass Eye’s outtakes show the brutal TV comedy was the tip of an iceberg - The Guardian',\n",
|
||||
" \"Vladimir Putin threatens civilians to break Mariupol's spirit - The Times\",\n",
|
||||
" 'Shell U-turn on Cambo oilfield would threaten green targets, say campaigners - The Guardian',\n",
|
||||
" 'St Helens dog attack: Girl aged 17 months killed at home - BBC',\n",
|
||||
" \"PlayStation to buy 'Assassin's Creed' veteran Jade Raymond's Haven Studios - NME\",\n",
|
||||
" '‘Clear sign’ Putin considering using chemical weapons in Ukraine, claims President Biden - The Independent',\n",
|
||||
" 'NASA confirms there are 5,000 planets outside our solar system - Daily Mail',\n",
|
||||
" 'Nintendo Switch finally has folders • Eurogamer.net - Eurogamer.net',\n",
|
||||
" 'FA to “find a solution” as Liverpool fan group blasts “shambolic” Wembley travel - This Is Anfield',\n",
|
||||
" 'Manchester United transfer news LIVE Erik ten Hag latest and Man Utd manager updates - Manchester Evening News',\n",
|
||||
" 'Inflation raises cost of UK government borrowing in February; crude oil up again – business live - The Guardian',\n",
|
||||
" 'Alexei Navalny: Kremlin critic found guilty of large-scale fraud and contempt of court by Russian court - Sky News',\n",
|
||||
" \"UK prepares to nationalize Russia natural gas giant Gazprom's retail unit - Business Insider\",\n",
|
||||
" 'Zaghari-Ratcliffe: Hunt calls for inquiry into delay over Iran debt payment - The Guardian']"
|
||||
]
|
||||
},
|
||||
"execution_count": 21,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"all_titles"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"首先，我們希望能夠從新聞標題中提取名詞。我們將使用 `TextBlob` 庫來完成這項工作，這可以簡化許多像這樣的典型自然語言處理任務。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 15,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Requirement already satisfied: textblob in c:\\winapp\\miniconda3\\lib\\site-packages (0.17.1)\n",
|
||||
"Requirement already satisfied: nltk>=3.1 in c:\\winapp\\miniconda3\\lib\\site-packages (from textblob) (3.5)\n",
|
||||
"Requirement already satisfied: joblib in c:\\winapp\\miniconda3\\lib\\site-packages (from nltk>=3.1->textblob) (1.0.1)\n",
|
||||
"Requirement already satisfied: regex in c:\\winapp\\miniconda3\\lib\\site-packages (from nltk>=3.1->textblob) (2021.11.10)\n",
|
||||
"Requirement already satisfied: tqdm in c:\\winapp\\miniconda3\\lib\\site-packages (from nltk>=3.1->textblob) (4.61.2)\n",
|
||||
"Requirement already satisfied: click in c:\\winapp\\miniconda3\\lib\\site-packages (from nltk>=3.1->textblob) (8.0.3)\n",
|
||||
"Requirement already satisfied: colorama in c:\\winapp\\miniconda3\\lib\\site-packages (from click->nltk>=3.1->textblob) (0.4.4)\n",
|
||||
"Finished.\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"name": "stderr",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"[nltk_data] Downloading package brown to\n",
|
||||
"[nltk_data] C:\\Users\\dmitryso\\AppData\\Roaming\\nltk_data...\n",
|
||||
"[nltk_data] Package brown is already up-to-date!\n",
|
||||
"[nltk_data] Downloading package punkt to\n",
|
||||
"[nltk_data] C:\\Users\\dmitryso\\AppData\\Roaming\\nltk_data...\n",
|
||||
"[nltk_data] Package punkt is already up-to-date!\n",
|
||||
"[nltk_data] Downloading package wordnet to\n",
|
||||
"[nltk_data] C:\\Users\\dmitryso\\AppData\\Roaming\\nltk_data...\n",
|
||||
"[nltk_data] Package wordnet is already up-to-date!\n",
|
||||
"[nltk_data] Downloading package averaged_perceptron_tagger to\n",
|
||||
"[nltk_data] C:\\Users\\dmitryso\\AppData\\Roaming\\nltk_data...\n",
|
||||
"[nltk_data] Package averaged_perceptron_tagger is already up-to-\n",
|
||||
"[nltk_data] date!\n",
|
||||
"[nltk_data] Downloading package conll2000 to\n",
|
||||
"[nltk_data] C:\\Users\\dmitryso\\AppData\\Roaming\\nltk_data...\n",
|
||||
"[nltk_data] Package conll2000 is already up-to-date!\n",
|
||||
"[nltk_data] Downloading package movie_reviews to\n",
|
||||
"[nltk_data] C:\\Users\\dmitryso\\AppData\\Roaming\\nltk_data...\n",
|
||||
"[nltk_data] Package movie_reviews is already up-to-date!\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"import sys\n",
|
||||
"!{sys.executable} -m pip install textblob\n",
|
||||
"!{sys.executable} -m textblob.download_corpora\n",
|
||||
"from textblob import TextBlob"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 22,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"{'covid-19 live updates': 1,\n",
|
||||
" 'vaccines': 1,\n",
|
||||
" 'boosters': 1,\n",
|
||||
" 'york': 4,\n",
|
||||
" 'ukrainians flee mariupol': 1,\n",
|
||||
" 'forces push': 1,\n",
|
||||
" 'port city': 1,\n",
|
||||
" 'wall street journal': 3,\n",
|
||||
" 'bond yields': 1,\n",
|
||||
" 'futures rise': 1,\n",
|
||||
" 'powell says fed': 1,\n",
|
||||
" 'ready': 1,\n",
|
||||
" 'be': 1,\n",
|
||||
" 'aggressive': 1,\n",
|
||||
" 'putin': 3,\n",
|
||||
" 'alexei navalny': 2,\n",
|
||||
" 'russian': 2,\n",
|
||||
" 'supreme court nominee': 1,\n",
|
||||
" 'ketanji brown jackson': 1,\n",
|
||||
" \"confirmation hearing 's\": 1,\n",
|
||||
" 'cnn': 1,\n",
|
||||
" 'swedish': 1,\n",
|
||||
" 'high school': 1,\n",
|
||||
" 'abc': 1,\n",
|
||||
" 'clues': 1,\n",
|
||||
" 'covid-19': 1,\n",
|
||||
" '’ s': 2,\n",
|
||||
" 'moves': 1,\n",
|
||||
" 'sewers': 1,\n",
|
||||
" 'roll dice': 1,\n",
|
||||
" 'jackson': 1,\n",
|
||||
" 'decisions |': 1,\n",
|
||||
" 'thehill': 1,\n",
|
||||
" 'clear': 2,\n",
|
||||
" 'chemical weapons': 2,\n",
|
||||
" 'ukraine': 3,\n",
|
||||
" 'claims president': 2,\n",
|
||||
" 'biden': 2,\n",
|
||||
" 'nasa': 2,\n",
|
||||
" 'solar system': 2,\n",
|
||||
" 'daily mail': 3,\n",
|
||||
" 'us stocks': 1,\n",
|
||||
" 'fed chair powell': 1,\n",
|
||||
" \"'s remarks\": 1,\n",
|
||||
" 'fox': 1,\n",
|
||||
" \"'we 've\": 1,\n",
|
||||
" 'tests': 1,\n",
|
||||
" 'covid': 1,\n",
|
||||
" 'politico': 1,\n",
|
||||
" 'duchess': 1,\n",
|
||||
" 'cambridge': 1,\n",
|
||||
" 'swaps khaki jungle gear': 1,\n",
|
||||
" 'vampire': 1,\n",
|
||||
" 'wife': 1,\n",
|
||||
" 'belize': 1,\n",
|
||||
" 'china': 2,\n",
|
||||
" 'flight recorders': 1,\n",
|
||||
" 'plane crash': 1,\n",
|
||||
" 'reuters': 2,\n",
|
||||
" 'russian oligarch': 1,\n",
|
||||
" 'abramovich': 1,\n",
|
||||
" 'live': 1,\n",
|
||||
" 'russia': 2,\n",
|
||||
" 'stops talks': 1,\n",
|
||||
" 'japan': 1,\n",
|
||||
" 'español': 1,\n",
|
||||
" 'powers remain': 1,\n",
|
||||
" 'threats lurk': 1,\n",
|
||||
" 'set': 1,\n",
|
||||
" 'webb': 1,\n",
|
||||
" 'telescope begins multi-instrument alignment': 1,\n",
|
||||
" 'scitechdaily': 1,\n",
|
||||
" 'uconn': 1,\n",
|
||||
" 'ucf': 1,\n",
|
||||
" 'ncaa': 1,\n",
|
||||
" \"women 's tournament second-round highlights\": 1,\n",
|
||||
" 'march madness': 1,\n",
|
||||
" 'bucking republican trend': 1,\n",
|
||||
" 'indiana': 1,\n",
|
||||
" 'vetoes transgender': 1,\n",
|
||||
" 'bill': 1,\n",
|
||||
" 'maggie fox': 1,\n",
|
||||
" 'coronation': 1,\n",
|
||||
" 'shameless': 1,\n",
|
||||
" \"'sudden accident\": 1,\n",
|
||||
" 'mirror online': 1,\n",
|
||||
" 'mirror': 2,\n",
|
||||
" 'plane crash –': 1,\n",
|
||||
" 'search': 1,\n",
|
||||
" 'moment flight': 1,\n",
|
||||
" 'daniel morgan': 1,\n",
|
||||
" 'report condemns': 1,\n",
|
||||
" 'met': 1,\n",
|
||||
" 'guardian': 6,\n",
|
||||
" 'rishi sunak': 1,\n",
|
||||
" '’ s spring': 1,\n",
|
||||
" 'statement': 1,\n",
|
||||
" 'bbc.com': 1,\n",
|
||||
" 'uk': 3,\n",
|
||||
" 'ireland': 1,\n",
|
||||
" 'euro': 1,\n",
|
||||
" 'vladimir putin': 2,\n",
|
||||
" \"'s 'lover\": 1,\n",
|
||||
" 'brass eye': 1,\n",
|
||||
" '’ s outtakes': 1,\n",
|
||||
" 'brutal tv comedy': 1,\n",
|
||||
" 'threatens civilians': 1,\n",
|
||||
" 'mariupol': 1,\n",
|
||||
" \"'s spirit\": 1,\n",
|
||||
" 'shell u-turn': 1,\n",
|
||||
" 'cambo': 1,\n",
|
||||
" 'green targets': 1,\n",
|
||||
" 'st helens': 1,\n",
|
||||
" 'dog attack': 1,\n",
|
||||
" 'girl': 1,\n",
|
||||
" 'bbc': 1,\n",
|
||||
" 'playstation': 1,\n",
|
||||
" \"'assassin 's\": 1,\n",
|
||||
" 'creed': 1,\n",
|
||||
" 'jade raymond': 1,\n",
|
||||
" 'haven studios': 1,\n",
|
||||
" 'nme': 1,\n",
|
||||
" 'nintendo switch': 1,\n",
|
||||
" 'folders •': 1,\n",
|
||||
" 'eurogamer.net': 2,\n",
|
||||
" 'fa': 1,\n",
|
||||
" 'solution ”': 1,\n",
|
||||
" 'liverpool': 1,\n",
|
||||
" 'fan group blasts “ shambolic ”': 1,\n",
|
||||
" 'wembley': 1,\n",
|
||||
" 'anfield': 1,\n",
|
||||
" 'manchester': 1,\n",
|
||||
" 'live erik': 1,\n",
|
||||
" 'hag': 1,\n",
|
||||
" 'utd': 1,\n",
|
||||
" 'manager updates': 1,\n",
|
||||
" 'manchester evening': 1,\n",
|
||||
" 'inflation': 1,\n",
|
||||
" 'government borrowing': 1,\n",
|
||||
" 'february': 1,\n",
|
||||
" 'crude oil': 1,\n",
|
||||
" '– business': 1,\n",
|
||||
" 'kremlin': 1,\n",
|
||||
" 'large-scale fraud': 1,\n",
|
||||
" 'sky': 1,\n",
|
||||
" 'natural gas': 1,\n",
|
||||
" 'gazprom': 1,\n",
|
||||
" 'retail unit': 1,\n",
|
||||
" 'insider': 1,\n",
|
||||
" 'zaghari-ratcliffe': 1,\n",
|
||||
" 'hunt': 1,\n",
|
||||
" 'iran': 1,\n",
|
||||
" 'debt payment': 1}"
|
||||
]
|
||||
},
|
||||
"execution_count": 22,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"w = {}\n",
|
||||
"for x in all_titles:\n",
|
||||
" for n in TextBlob(x).noun_phrases:\n",
|
||||
" if n in w:\n",
|
||||
" w[n].append(x)\n",
|
||||
" else:\n",
|
||||
" w[n]=[x]\n",
|
||||
"{ x:len(w[x]) for x in w.keys()}"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們可以看到，名詞並未給我們帶來大的主題群組。讓我們用從概念圖中獲得的更一般的術語來替換名詞。這將需要一些時間，因為我們正在為每個名詞短語進行 REST 調用。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 23,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"w = {}\n",
|
||||
"for x in all_titles:\n",
|
||||
" for noun in TextBlob(x).noun_phrases:\n",
|
||||
" terms = query(noun.replace(' ','%20'))\n",
|
||||
" for term in [u for u in terms.keys() if terms[u]>0.1]:\n",
|
||||
" if term in w:\n",
|
||||
" w[term].append(x)\n",
|
||||
" else:\n",
|
||||
" w[term]=[x]"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 24,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"{'city': 9,\n",
|
||||
" 'brand': 4,\n",
|
||||
" 'place': 9,\n",
|
||||
" 'town': 4,\n",
|
||||
" 'factor': 4,\n",
|
||||
" 'film': 4,\n",
|
||||
" 'nation': 11,\n",
|
||||
" 'state': 5,\n",
|
||||
" 'person': 4,\n",
|
||||
" 'organization': 5,\n",
|
||||
" 'publication': 10,\n",
|
||||
" 'market': 5,\n",
|
||||
" 'economy': 4,\n",
|
||||
" 'company': 6,\n",
|
||||
" 'newspaper': 6,\n",
|
||||
" 'relationship': 6}"
|
||||
]
|
||||
},
|
||||
"execution_count": 24,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"{ x:len(w[x]) for x in w.keys() if len(w[x])>3}"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 27,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"\n",
|
||||
"ECONOMY:\n",
|
||||
"China searches for victims, flight recorders after first plane crash in 12 years - Reuters\n",
|
||||
"Live updates: Russia stops talks with Japan over sanctions - The Associated Press - en Español\n",
|
||||
"China plane crash – live: Search for survivors continues as witness describes moment flight fell from sky - The Independent\n",
|
||||
"UK prepares to nationalize Russia natural gas giant Gazprom's retail unit - Business Insider\n",
|
||||
"\n",
|
||||
"NATION:\n",
|
||||
"‘Clear sign’ Putin considering using chemical weapons in Ukraine, claims President Biden - The Independent\n",
|
||||
"Duchess of Cambridge swaps khaki jungle gear for Vampire's Wife dress on Belize trip - Daily Mail\n",
|
||||
"China searches for victims, flight recorders after first plane crash in 12 years - Reuters\n",
|
||||
"Live updates: Russia stops talks with Japan over sanctions - The Associated Press - en Español\n",
|
||||
"Live updates: Russia stops talks with Japan over sanctions - The Associated Press - en Español\n",
|
||||
"China plane crash – live: Search for survivors continues as witness describes moment flight fell from sky - The Independent\n",
|
||||
"UK and Republic of Ireland in line to host Euro 2028 after no one else bids - The Guardian\n",
|
||||
"Friends beg Vladimir Putin's 'lover' to persuade him to end Ukraine invasion - The Mirror\n",
|
||||
"‘Clear sign’ Putin considering using chemical weapons in Ukraine, claims President Biden - The Independent\n",
|
||||
"UK prepares to nationalize Russia natural gas giant Gazprom's retail unit - Business Insider\n",
|
||||
"Zaghari-Ratcliffe: Hunt calls for inquiry into delay over Iran debt payment - The Guardian\n",
|
||||
"\n",
|
||||
"PERSON:\n",
|
||||
"‘Clear sign’ Putin considering using chemical weapons in Ukraine, claims President Biden - The Independent\n",
|
||||
"Duchess of Cambridge swaps khaki jungle gear for Vampire's Wife dress on Belize trip - Daily Mail\n",
|
||||
"Second superyacht linked to Russian oligarch Abramovich docks in Turkey - Reuters\n",
|
||||
"‘Clear sign’ Putin considering using chemical weapons in Ukraine, claims President Biden - The Independent\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"print('\\nECONOMY:\\n'+'\\n'.join(w['economy']))\n",
|
||||
"print('\\nNATION:\\n'+'\\n'.join(w['nation']))\n",
|
||||
"print('\\nPERSON:\\n'+'\\n'.join(w['person']))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件已使用人工智能翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。應以原文文件作為權威來源。對於關鍵資訊，建議尋求專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或誤釋不承擔責任。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"kernelspec": {
|
||||
"display_name": "Python 3.7.4 64-bit (conda)",
|
||||
"metadata": {
|
||||
"interpreter": {
|
||||
"hash": "86193a1ab0ba47eac1c69c1756090baa3b420b3eea7d4aafab8b85f8b312f0c5"
|
||||
}
|
||||
},
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.9.5"
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "4087f998407d06ceb2947016ba4605d0",
|
||||
"translation_date": "2025-08-31T10:06:49+00:00",
|
||||
"source_file": "lessons/2-Symbolic/MSConceptGraph.ipynb",
|
||||
"language_code": "hk"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 2
|
||||
}
|
||||
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
|
|
@ -1,31 +1,33 @@
|
|||
<!--
|
||||
CO_OP_TRANSLATOR_METADATA:
|
||||
{
|
||||
"original_hash": "7336583e4630220c835335da640016db",
|
||||
"translation_date": "2025-08-24T22:10:58+00:00",
|
||||
"original_hash": "ba5d1eb353d20d3e7181066b3c424b99",
|
||||
"translation_date": "2025-08-31T09:20:54+00:00",
|
||||
"source_file": "lessons/3-NeuralNetworks/03-Perceptron/lab/README.md",
|
||||
"language_code": "hk"
|
||||
}
|
||||
-->
|
||||
# 使用感知器進行多類別分類
|
||||
|
||||
來自 [AI for Beginners Curriculum](https://github.com/microsoft/ai-for-beginners) 的實驗作業。
|
||||
來自 [AI for Beginners Curriculum](https://github.com/microsoft/ai-for-beginners) 的實驗課題。
|
||||
|
||||
## 任務
|
||||
|
||||
使用我們在本課中為 MNIST 手寫數字進行二元分類所開發的程式碼，創建一個多類別分類器，能夠識別任何數字。計算訓練集和測試集的分類準確率，並輸出混淆矩陣。
|
||||
利用我們在本課中為 MNIST 手寫數字進行二元分類所開發的程式碼，創建一個多類別分類器，能夠識別任意數字。計算訓練集和測試集的分類準確率，並輸出混淆矩陣。
|
||||
|
||||
## 提示
|
||||
|
||||
1. 對於每個數字，創建一個二元分類器的數據集，將其設為「該數字 vs. 其他所有數字」
|
||||
1. 訓練 10 個不同的感知器進行二元分類（每個數字一個感知器）
|
||||
1. 定義一個函數來分類輸入的數字
|
||||
1. 對於每個數字，創建一個二元分類的數據集，將「該數字」與「所有其他數字」進行分類。
|
||||
1. 訓練 10 個不同的感知器進行二元分類（每個數字對應一個感知器）。
|
||||
1. 定義一個函數來分類輸入的數字。
|
||||
|
||||
> **提示**：如果我們將所有 10 個感知器的權重組合成一個矩陣，我們應該能夠通過一次矩陣乘法將所有 10 個感知器應用到輸入數字上。最有可能的數字可以通過對輸出應用 `argmax` 操作來找到。
|
||||
> **提示**: 如果我們將所有 10 個感知器的權重組合成一個矩陣，我們應該能夠通過一次矩陣乘法將所有 10 個感知器應用到輸入數字上。最可能的數字可以通過對輸出應用 `argmax` 操作來找到。
|
||||
|
||||
## 起始筆記本
|
||||
|
||||
通過打開 [PerceptronMultiClass.ipynb](../../../../../../lessons/3-NeuralNetworks/03-Perceptron/lab/PerceptronMultiClass.ipynb) 開始實驗。
|
||||
通過打開 [PerceptronMultiClass.ipynb](PerceptronMultiClass.ipynb) 開始實驗課題。
|
||||
|
||||
---
|
||||
|
||||
**免責聲明**：
|
||||
本文件已使用人工智能翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。應以原文文件作為權威來源。對於關鍵資訊，建議尋求專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或誤釋不承擔責任。
|
||||
本文件已使用人工智能翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。儘管我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。原始文件的母語版本應被視為權威來源。對於重要信息，建議使用專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或錯誤解釋概不負責。
|
||||
File diff suppressed because one or more lines are too long
|
|
@ -0,0 +1,183 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 使用我們自己的框架進行 MNIST 數字分類\n",
|
||||
"\n",
|
||||
"來自 [AI 初學者課程](https://github.com/microsoft/ai-for-beginners) 的實驗作業。\n",
|
||||
"\n",
|
||||
"### 讀取數據集\n",
|
||||
"\n",
|
||||
"此程式碼會從網絡上的存儲庫下載數據集。你也可以手動從 AI 課程存儲庫的 `/data` 目錄中複製數據集。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {
|
||||
"tags": []
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stderr",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
" % Total % Received % Xferd Average Speed Time Time Time Current\n",
|
||||
" Dload Upload Total Spent Left Speed\n",
|
||||
"\n",
|
||||
" 0 0 0 0 0 0 0 0 --:--:-- --:--:-- --:--:-- 0\n",
|
||||
"100 9.9M 100 9.9M 0 0 9.9M 0 0:00:01 --:--:-- 0:00:01 15.8M\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"!rm *.pkl\n",
|
||||
"!wget https://raw.githubusercontent.com/microsoft/AI-For-Beginners/main/data/mnist.pkl.gz\n",
|
||||
"!gzip -d mnist.pkl.gz"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import pickle\n",
|
||||
"with open('mnist.pkl','rb') as f:\n",
|
||||
" MNIST = pickle.load(f)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"labels = MNIST['Train']['Labels']\n",
|
||||
"data = MNIST['Train']['Features']"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"讓我們看看我們擁有的數據形狀：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"(42000, 784)"
|
||||
]
|
||||
},
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"data.shape"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"### 分割數據\n",
|
||||
"\n",
|
||||
"我們將使用 Scikit Learn 將數據分為訓練集和測試集：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 6,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Train samples: 33600, test samples: 8400\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"from sklearn.model_selection import train_test_split\n",
|
||||
"\n",
|
||||
"features_train, features_test, labels_train, labels_test = train_test_split(data,labels,test_size=0.2)\n",
|
||||
"\n",
|
||||
"print(f\"Train samples: {len(features_train)}, test samples: {len(features_test)}\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"### 指示\n",
|
||||
"\n",
|
||||
"1. 從課堂中取得框架代碼，並將其貼到此筆記本中，或（更好）貼到一個獨立的 Python 模組中\n",
|
||||
"1. 定義並訓練單層感知器，觀察訓練和驗證的準確率在訓練過程中的變化\n",
|
||||
"1. 嘗試理解是否出現過度擬合，並調整層的參數以改善準確率\n",
|
||||
"1. 重複上述步驟，訓練兩層和三層感知器。嘗試在層之間使用不同的激活函數進行實驗\n",
|
||||
"1. 嘗試回答以下問題：\n",
|
||||
" - 層間的激活函數是否影響網絡性能？\n",
|
||||
" - 這個任務是否需要兩層或三層的網絡？\n",
|
||||
" - 在訓練網絡時是否遇到任何問題？尤其是當層數增加時。\n",
|
||||
" - 網絡的權重在訓練過程中如何變化？你可以繪製權重的最大絕對值與訓練輪次的關係圖來理解這種關係。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": []
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件已使用人工智能翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。應以原文文件作為權威來源。如涉及關鍵資訊，建議尋求專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或誤釋概不負責。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"kernelspec": {
|
||||
"display_name": "Python 3.7.4 64-bit (conda)",
|
||||
"metadata": {
|
||||
"interpreter": {
|
||||
"hash": "86193a1ab0ba47eac1c69c1756090baa3b420b3eea7d4aafab8b85f8b312f0c5"
|
||||
}
|
||||
},
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.9.5"
|
||||
},
|
||||
"orig_nbformat": 2,
|
||||
"coopTranslator": {
|
||||
"original_hash": "6fa055f484eb5d6bdf41166a356d3abf",
|
||||
"translation_date": "2025-08-31T10:13:14+00:00",
|
||||
"source_file": "lessons/3-NeuralNetworks/04-OwnFramework/lab/MyFW_MNIST.ipynb",
|
||||
"language_code": "hk"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 2
|
||||
}
|
||||
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
|
|
@ -0,0 +1,102 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"**你的目標**是使用光流技術來判斷影片中哪些部分包含向上、向下、向左或向右的移動。\n",
|
||||
"\n",
|
||||
"首先按照課堂中所述的方法提取影片的幀：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"# Code here"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在，按照講座中所描述的計算密集光流幀，並將密集光流轉換為極座標：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"# Code here"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"為每個光流幀建立方向的直方圖。直方圖顯示有多少向量落入特定的區間，並應該將幀中不同方向的移動分開。\n",
|
||||
"\n",
|
||||
"> 你可能也需要將所有大小低於某個閾值的向量設為零。這樣可以去除影片中一些小的額外移動，例如眼睛和頭部。\n",
|
||||
"\n",
|
||||
"繪製一些幀的直方圖。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"# Code here"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"看著直方圖，應該可以很直接地判斷移動的方向。你需要選擇那些對應於上/下/左/右方向的區間，並且超過某個閾值的區間。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"# Code here"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"恭喜！如果你已完成以上所有步驟，你已完成此實驗！\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件已使用人工智能翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。儘管我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。原始語言的文件應被視為權威來源。對於重要信息，建議使用專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或錯誤解釋不承擔責任。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"language_info": {
|
||||
"name": "python"
|
||||
},
|
||||
"orig_nbformat": 4,
|
||||
"coopTranslator": {
|
||||
"original_hash": "153d9e417e079bf62f8f693002d0deaf",
|
||||
"translation_date": "2025-08-31T09:49:51+00:00",
|
||||
"source_file": "lessons/4-ComputerVision/06-IntroCV/lab/MovementDetection.ipynb",
|
||||
"language_code": "hk"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 2
|
||||
}
|
||||
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
|
|
@ -0,0 +1,577 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 文本分類任務\n",
|
||||
"\n",
|
||||
"如前所述，我們將專注於基於 **AG_NEWS** 數據集的簡單文本分類任務，目的是將新聞標題分類為以下四個類別之一：國際、體育、商業和科學/技術。\n",
|
||||
"\n",
|
||||
"## 數據集\n",
|
||||
"\n",
|
||||
"此數據集已內建於 [`torchtext`](https://github.com/pytorch/text) 模組中，因此我們可以輕鬆存取它。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 1,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import torch\n",
|
||||
"import torchtext\n",
|
||||
"import os\n",
|
||||
"import collections\n",
|
||||
"os.makedirs('./data',exist_ok=True)\n",
|
||||
"train_dataset, test_dataset = torchtext.datasets.AG_NEWS(root='./data')\n",
|
||||
"classes = ['World', 'Sports', 'Business', 'Sci/Tech']"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"在這裡，`train_dataset` 和 `test_dataset` 包含分別返回標籤（類別編號）和文本的集合，例如：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 2,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"(3,\n",
|
||||
" \"Wall St. Bears Claw Back Into the Black (Reuters) Reuters - Short-sellers, Wall Street's dwindling\\\\band of ultra-cynics, are seeing green again.\")"
|
||||
]
|
||||
},
|
||||
"execution_count": 2,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"list(train_dataset)[0]"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"所以，讓我們列印出資料集中的前10個新標題：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"**Sci/Tech** -> Wall St. Bears Claw Back Into the Black (Reuters) Reuters - Short-sellers, Wall Street's dwindling\\band of ultra-cynics, are seeing green again.\n",
|
||||
"**Sci/Tech** -> Carlyle Looks Toward Commercial Aerospace (Reuters) Reuters - Private investment firm Carlyle Group,\\which has a reputation for making well-timed and occasionally\\controversial plays in the defense industry, has quietly placed\\its bets on another part of the market.\n",
|
||||
"**Sci/Tech** -> Oil and Economy Cloud Stocks' Outlook (Reuters) Reuters - Soaring crude prices plus worries\\about the economy and the outlook for earnings are expected to\\hang over the stock market next week during the depth of the\\summer doldrums.\n",
|
||||
"**Sci/Tech** -> Iraq Halts Oil Exports from Main Southern Pipeline (Reuters) Reuters - Authorities have halted oil export\\flows from the main pipeline in southern Iraq after\\intelligence showed a rebel militia could strike\\infrastructure, an oil official said on Saturday.\n",
|
||||
"**Sci/Tech** -> Oil prices soar to all-time record, posing new menace to US economy (AFP) AFP - Tearaway world oil prices, toppling records and straining wallets, present a new economic menace barely three months before the US presidential elections.\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"for i,x in zip(range(5),train_dataset):\n",
|
||||
" print(f\"**{classes[x[0]]}** -> {x[1]}\")\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"因為數據集是迭代器，如果我們想多次使用數據，我們需要將其轉換為列表：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"train_dataset, test_dataset = torchtext.datasets.AG_NEWS(root='./data')\n",
|
||||
"train_dataset = list(train_dataset)\n",
|
||||
"test_dataset = list(test_dataset)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 分詞\n",
|
||||
"\n",
|
||||
"現在我們需要將文本轉換成**數字**，以便表示為張量。如果我們想要詞級表示，需要完成以下兩件事：\n",
|
||||
"* 使用**分詞器**將文本拆分成**詞元**\n",
|
||||
"* 建立這些詞元的**詞彙表**。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"['he', 'said', 'hello']"
|
||||
]
|
||||
},
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"tokenizer = torchtext.data.utils.get_tokenizer('basic_english')\n",
|
||||
"tokenizer('He said: hello')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"counter = collections.Counter()\n",
|
||||
"for (label, line) in train_dataset:\n",
|
||||
" counter.update(tokenizer(line))\n",
|
||||
"vocab = torchtext.vocab.vocab(counter, min_freq=1)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"使用詞彙表，我們可以輕鬆地將標記化的字串編碼為一組數字：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 19,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Vocab size if 95810\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"[599, 3279, 97, 1220, 329, 225, 7368]"
|
||||
]
|
||||
},
|
||||
"execution_count": 19,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"vocab_size = len(vocab)\n",
|
||||
"print(f\"Vocab size if {vocab_size}\")\n",
|
||||
"\n",
|
||||
"stoi = vocab.get_stoi() # dict to convert tokens to indices\n",
|
||||
"\n",
|
||||
"def encode(x):\n",
|
||||
" return [stoi[s] for s in tokenizer(x)]\n",
|
||||
"\n",
|
||||
"encode('I love to play with my words')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 詞袋文字表示法\n",
|
||||
"\n",
|
||||
"由於文字代表意義，有時我們可以僅透過查看個別的文字來理解文本的含義，而不需要考慮它們在句子中的順序。例如，在分類新聞時，像 *天氣*、*雪* 這些詞可能表明是 *天氣預報*，而像 *股票*、*美元* 則可能屬於 *財經新聞*。\n",
|
||||
"\n",
|
||||
"**詞袋** (BoW) 向量表示法是最常用的傳統向量表示法之一。每個文字都與向量索引相關聯，向量元素包含某個文字在特定文件中出現的次數。\n",
|
||||
"\n",
|
||||
" \n",
|
||||
"\n",
|
||||
"> **注意**：你也可以將 BoW 理解為文本中每個文字的獨立一熱編碼向量的總和。\n",
|
||||
"\n",
|
||||
"以下是一個使用 Scikit Learn Python 庫生成詞袋表示法的示例：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"array([[1, 1, 0, 2, 0, 0, 0, 0, 0]], dtype=int64)"
|
||||
]
|
||||
},
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"from sklearn.feature_extraction.text import CountVectorizer\n",
|
||||
"vectorizer = CountVectorizer()\n",
|
||||
"corpus = [\n",
|
||||
" 'I like hot dogs.',\n",
|
||||
" 'The dog ran fast.',\n",
|
||||
" 'Its hot outside.',\n",
|
||||
" ]\n",
|
||||
"vectorizer.fit_transform(corpus)\n",
|
||||
"vectorizer.transform(['My dog likes hot dogs on a hot day.']).toarray()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"要從我們的 AG_NEWS 數據集的向量表示計算詞袋向量，可以使用以下函數：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 20,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"tensor([2., 1., 2., ..., 0., 0., 0.])\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"vocab_size = len(vocab)\n",
|
||||
"\n",
|
||||
"def to_bow(text,bow_vocab_size=vocab_size):\n",
|
||||
" res = torch.zeros(bow_vocab_size,dtype=torch.float32)\n",
|
||||
" for i in encode(text):\n",
|
||||
" if i<bow_vocab_size:\n",
|
||||
" res[i] += 1\n",
|
||||
" return res\n",
|
||||
"\n",
|
||||
"print(to_bow(train_dataset[0][1]))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"> **注意：** 這裡我們使用全域變數 `vocab_size` 來指定詞彙表的預設大小。由於詞彙表的大小通常相當大，我們可以將詞彙表的大小限制為最常見的詞彙。嘗試降低 `vocab_size` 的值並運行以下程式碼，觀察它如何影響準確性。你應該預期會有一些準確性的下降，但不會太劇烈，以換取更高的效能。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 訓練 BoW 分類器\n",
|
||||
"\n",
|
||||
"現在我們已經學會如何建立文字的詞袋表示法，接下來讓我們在此基礎上訓練一個分類器。首先，我們需要將數據集轉換為適合訓練的格式，將所有位置向量表示法轉換為詞袋表示法。這可以通過將 `bowify` 函數作為標準 torch `DataLoader` 的 `collate_fn` 參數來實現：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 21,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"from torch.utils.data import DataLoader\n",
|
||||
"import numpy as np \n",
|
||||
"\n",
|
||||
"# this collate function gets list of batch_size tuples, and needs to \n",
|
||||
"# return a pair of label-feature tensors for the whole minibatch\n",
|
||||
"def bowify(b):\n",
|
||||
" return (\n",
|
||||
" torch.LongTensor([t[0]-1 for t in b]),\n",
|
||||
" torch.stack([to_bow(t[1]) for t in b])\n",
|
||||
" )\n",
|
||||
"\n",
|
||||
"train_loader = DataLoader(train_dataset, batch_size=16, collate_fn=bowify, shuffle=True)\n",
|
||||
"test_loader = DataLoader(test_dataset, batch_size=16, collate_fn=bowify, shuffle=True)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在讓我們定義一個簡單的分類器神經網絡，其中包含一個線性層。輸入向量的大小等於 `vocab_size`，輸出大小對應於類別數量（4）。由於我們正在解決分類任務，最終的激活函數是 `LogSoftmax()`。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 22,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"net = torch.nn.Sequential(torch.nn.Linear(vocab_size,4),torch.nn.LogSoftmax(dim=1))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在我們將定義標準的 PyTorch 訓練循環。由於我們的數據集相當大，為了教學目的，我們只會訓練一個 epoch，有時甚至少於一個 epoch（通過指定 `epoch_size` 參數可以限制訓練）。我們還會在訓練過程中報告累積的訓練準確率；報告的頻率是通過 `report_freq` 參數指定的。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 24,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"def train_epoch(net,dataloader,lr=0.01,optimizer=None,loss_fn = torch.nn.NLLLoss(),epoch_size=None, report_freq=200):\n",
|
||||
" optimizer = optimizer or torch.optim.Adam(net.parameters(),lr=lr)\n",
|
||||
" net.train()\n",
|
||||
" total_loss,acc,count,i = 0,0,0,0\n",
|
||||
" for labels,features in dataloader:\n",
|
||||
" optimizer.zero_grad()\n",
|
||||
" out = net(features)\n",
|
||||
" loss = loss_fn(out,labels) #cross_entropy(out,labels)\n",
|
||||
" loss.backward()\n",
|
||||
" optimizer.step()\n",
|
||||
" total_loss+=loss\n",
|
||||
" _,predicted = torch.max(out,1)\n",
|
||||
" acc+=(predicted==labels).sum()\n",
|
||||
" count+=len(labels)\n",
|
||||
" i+=1\n",
|
||||
" if i%report_freq==0:\n",
|
||||
" print(f\"{count}: acc={acc.item()/count}\")\n",
|
||||
" if epoch_size and count>epoch_size:\n",
|
||||
" break\n",
|
||||
" return total_loss.item()/count, acc.item()/count"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 25,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"3200: acc=0.8028125\n",
|
||||
"6400: acc=0.8371875\n",
|
||||
"9600: acc=0.8534375\n",
|
||||
"12800: acc=0.85765625\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"(0.026090790722161722, 0.8620069296375267)"
|
||||
]
|
||||
},
|
||||
"execution_count": 25,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"train_epoch(net,train_loader,epoch_size=15000)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## BiGrams、TriGrams 和 N-Grams\n",
|
||||
"\n",
|
||||
"袋裝詞語（Bag of Words）方法的一個限制是，有些詞語是多詞組成的表達，例如「hot dog」這個詞語的意思與「hot」和「dog」在其他語境中的意思完全不同。如果我們總是用相同的向量來表示「hot」和「dog」，可能會讓模型感到混淆。\n",
|
||||
"\n",
|
||||
"為了解決這個問題，**N-gram 表示法**經常被用於文件分類的方法中，其中每個單詞、雙詞或三詞的出現頻率是訓練分類器的一個有用特徵。例如，在雙詞組（bigram）表示法中，除了原本的單詞外，我們還會將所有的詞對加入詞彙表中。\n",
|
||||
"\n",
|
||||
"以下是一個使用 Scikit Learn 生成雙詞組袋裝詞語表示法的例子：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 26,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Vocabulary:\n",
|
||||
" {'i': 7, 'like': 11, 'hot': 4, 'dogs': 2, 'i like': 8, 'like hot': 12, 'hot dogs': 5, 'the': 16, 'dog': 0, 'ran': 14, 'fast': 3, 'the dog': 17, 'dog ran': 1, 'ran fast': 15, 'its': 9, 'outside': 13, 'its hot': 10, 'hot outside': 6}\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"array([[1, 0, 1, 0, 2, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]],\n",
|
||||
" dtype=int64)"
|
||||
]
|
||||
},
|
||||
"execution_count": 26,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"bigram_vectorizer = CountVectorizer(ngram_range=(1, 2), token_pattern=r'\\b\\w+\\b', min_df=1)\n",
|
||||
"corpus = [\n",
|
||||
" 'I like hot dogs.',\n",
|
||||
" 'The dog ran fast.',\n",
|
||||
" 'Its hot outside.',\n",
|
||||
" ]\n",
|
||||
"bigram_vectorizer.fit_transform(corpus)\n",
|
||||
"print(\"Vocabulary:\\n\",bigram_vectorizer.vocabulary_)\n",
|
||||
"bigram_vectorizer.transform(['My dog likes hot dogs on a hot day.']).toarray()\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"N-gram 方法的主要缺點是詞彙量會迅速增長。在實際應用中，我們需要將 N-gram 表示法與一些降維技術結合使用，例如 *嵌入*，我們會在下一單元中討論。\n",
|
||||
"\n",
|
||||
"要在我們的 **AG News** 數據集中使用 N-gram 表示法，我們需要建立專門的 ngram 詞彙：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 27,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Bigram vocabulary length = 1308842\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"counter = collections.Counter()\n",
|
||||
"for (label, line) in train_dataset:\n",
|
||||
" l = tokenizer(line)\n",
|
||||
" counter.update(torchtext.data.utils.ngrams_iterator(l,ngrams=2))\n",
|
||||
" \n",
|
||||
"bi_vocab = torchtext.vocab.vocab(counter, min_freq=1)\n",
|
||||
"\n",
|
||||
"print(\"Bigram vocabulary length = \",len(bi_vocab))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們可以使用上述相同的程式碼來訓練分類器，但這樣會非常耗費記憶體。在下一個單元中，我們將使用嵌入來訓練二元語法分類器。\n",
|
||||
"\n",
|
||||
"> **注意：** 你只能保留那些在文本中出現次數超過指定數量的 ngrams。這樣可以確保罕見的二元語法會被省略，並顯著減少維度。為此，將 `min_freq` 參數設置為更高的值，並觀察詞彙表長度的變化。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 詞頻-逆文檔頻率 TF-IDF\n",
|
||||
"\n",
|
||||
"在 BoW 表示法中，詞語的出現次數被均等地加權，無論詞語本身的特性如何。然而，很明顯像 *a*、*in* 等這些常見詞對分類的作用遠不如一些專業術語。事實上，在大多數 NLP 任務中，有些詞語比其他詞語更具相關性。\n",
|
||||
"\n",
|
||||
"**TF-IDF** 代表 **詞頻–逆文檔頻率**。它是袋子模型（BoW）的變體，與使用二進制 0/1 值來表示詞語是否出現在文檔中的方法不同，TF-IDF 使用浮點值，該值與詞語在語料庫中的出現頻率相關。\n",
|
||||
"\n",
|
||||
"更正式地說，詞語 $i$ 在文檔 $j$ 中的權重 $w_{ij}$ 定義如下：\n",
|
||||
"$$\n",
|
||||
"w_{ij} = tf_{ij}\\times\\log({N\\over df_i})\n",
|
||||
"$$\n",
|
||||
"其中：\n",
|
||||
"* $tf_{ij}$ 是詞語 $i$ 在文檔 $j$ 中的出現次數，即我們之前看到的 BoW 值\n",
|
||||
"* $N$ 是語料庫中的文檔總數\n",
|
||||
"* $df_i$ 是包含詞語 $i$ 的文檔數量\n",
|
||||
"\n",
|
||||
"TF-IDF 值 $w_{ij}$ 與詞語在文檔中出現的次數成正比，但會根據語料庫中包含該詞語的文檔數量進行調整，這有助於平衡某些詞語出現頻率較高的情況。例如，如果某個詞語出現在語料庫中的 *每一份* 文檔中，則 $df_i=N$，而 $w_{ij}=0$，這些詞語將被完全忽略。\n",
|
||||
"\n",
|
||||
"你可以使用 Scikit Learn 輕鬆地創建文本的 TF-IDF 向量化：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 28,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"array([[0.43381609, 0. , 0.43381609, 0. , 0.65985664,\n",
|
||||
" 0.43381609, 0. , 0. , 0. , 0. ,\n",
|
||||
" 0. , 0. , 0. , 0. , 0. ,\n",
|
||||
" 0. ]])"
|
||||
]
|
||||
},
|
||||
"execution_count": 28,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"from sklearn.feature_extraction.text import TfidfVectorizer\n",
|
||||
"vectorizer = TfidfVectorizer(ngram_range=(1,2))\n",
|
||||
"vectorizer.fit_transform(corpus)\n",
|
||||
"vectorizer.transform(['My dog likes hot dogs on a hot day.']).toarray()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 結論\n",
|
||||
"\n",
|
||||
"然而，儘管 TF-IDF 表示法為不同的詞語提供了頻率權重，但它無法表達詞語的意義或順序。正如著名語言學家 J. R. Firth 在1935年所說：「詞語的完整意義總是與上下文相關，任何脫離上下文的意義研究都不應被認真對待。」我們稍後會在課程中學習如何通過語言建模從文本中捕捉上下文信息。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件已使用人工智能翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。應以原文文件作為權威來源。對於關鍵資訊，建議尋求專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或誤釋不承擔責任。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"interpreter": {
|
||||
"hash": "16af2a8bbb083ea23e5e41c7f5787656b2ce26968575d8763f2c4b17f9cd711f"
|
||||
},
|
||||
"kernelspec": {
|
||||
"display_name": "Python 3.8.12 ('py38')",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.8.12"
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "7b9040985e748e4e2d4c689892456ad7",
|
||||
"translation_date": "2025-08-31T10:59:02+00:00",
|
||||
"source_file": "lessons/5-NLP/13-TextRep/TextRepresentationPyTorch.ipynb",
|
||||
"language_code": "hk"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 2
|
||||
}
|
||||
|
|
@ -0,0 +1,647 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 文本分類任務\n",
|
||||
"\n",
|
||||
"在這個模組中，我們將從一個簡單的文本分類任務開始，基於 **[AG_NEWS](http://www.di.unipi.it/~gulli/AG_corpus_of_news_articles.html)** 數據集：我們會將新聞標題分類為以下四個類別之一：世界、體育、商業和科技。\n",
|
||||
"\n",
|
||||
"## 數據集\n",
|
||||
"\n",
|
||||
"為了載入數據集，我們將使用 **[TensorFlow Datasets](https://www.tensorflow.org/datasets)** API。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 1,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import tensorflow as tf\n",
|
||||
"from tensorflow import keras\n",
|
||||
"import tensorflow_datasets as tfds\n",
|
||||
"\n",
|
||||
"# In this tutorial, we will be training a lot of models. In order to use GPU memory cautiously,\n",
|
||||
"# we will set tensorflow option to grow GPU memory allocation when required.\n",
|
||||
"physical_devices = tf.config.list_physical_devices('GPU') \n",
|
||||
"if len(physical_devices)>0:\n",
|
||||
" tf.config.experimental.set_memory_growth(physical_devices[0], True)\n",
|
||||
"\n",
|
||||
"dataset = tfds.load('ag_news_subset')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們現在可以分別使用 `dataset['train']` 和 `dataset['test']` 來訪問數據集的訓練和測試部分：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Length of train dataset = 120000\n",
|
||||
"Length of test dataset = 7600\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"ds_train = dataset['train']\n",
|
||||
"ds_test = dataset['test']\n",
|
||||
"\n",
|
||||
"print(f\"Length of train dataset = {len(ds_train)}\")\n",
|
||||
"print(f\"Length of test dataset = {len(ds_test)}\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"讓我們列印出資料集中前10個新的標題：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"3 (Sci/Tech) -> b'AMD Debuts Dual-Core Opteron Processor' b'AMD #39;s new dual-core Opteron chip is designed mainly for corporate computing applications, including databases, Web services, and financial transactions.'\n",
|
||||
"1 (Sports) -> b\"Wood's Suspension Upheld (Reuters)\" b'Reuters - Major League Baseball\\\\Monday announced a decision on the appeal filed by Chicago Cubs\\\\pitcher Kerry Wood regarding a suspension stemming from an\\\\incident earlier this season.'\n",
|
||||
"2 (Business) -> b'Bush reform may have blue states seeing red' b'President Bush #39;s quot;revenue-neutral quot; tax reform needs losers to balance its winners, and people claiming the federal deduction for state and local taxes may be in administration planners #39; sights, news reports say.'\n",
|
||||
"3 (Sci/Tech) -> b\"'Halt science decline in schools'\" b'Britain will run out of leading scientists unless science education is improved, says Professor Colin Pillinger.'\n",
|
||||
"1 (Sports) -> b'Gerrard leaves practice' b'London, England (Sports Network) - England midfielder Steven Gerrard injured his groin late in Thursday #39;s training session, but is hopeful he will be ready for Saturday #39;s World Cup qualifier against Austria.'\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"classes = ['World', 'Sports', 'Business', 'Sci/Tech']\n",
|
||||
"\n",
|
||||
"for i,x in zip(range(5),ds_train):\n",
|
||||
" print(f\"{x['label']} ({classes[x['label']]}) -> {x['title']} {x['description']}\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 文本向量化\n",
|
||||
"\n",
|
||||
"現在我們需要將文本轉換成可以表示為張量的**數字**。如果我們想要詞級別的表示，需要完成以下兩件事：\n",
|
||||
"\n",
|
||||
"* 使用**分詞器**將文本拆分為**詞元**。\n",
|
||||
"* 建立這些詞元的**詞彙表**。\n",
|
||||
"\n",
|
||||
"### 限制詞彙表大小\n",
|
||||
"\n",
|
||||
"在 AG News 數據集的例子中，詞彙表的大小相當大，超過 100k 個單詞。一般來說，我們不需要那些在文本中很少出現的單詞——只有少數句子會包含它們，而模型無法從中學習。因此，通過向向量化器構造函數傳遞參數，將詞彙表的大小限制為較小的數量是有意義的：\n",
|
||||
"\n",
|
||||
"這兩個步驟都可以使用 **TextVectorization** 層來處理。我們來實例化向量化器對象，然後調用 `adapt` 方法來遍歷所有文本並建立詞彙表：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"vocab_size = 50000\n",
|
||||
"vectorizer = keras.layers.experimental.preprocessing.TextVectorization(max_tokens=vocab_size)\n",
|
||||
"vectorizer.adapt(ds_train.take(500).map(lambda x: x['title']+' '+x['description']))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"> **注意** 我們只使用整個數據集的一部分來建立詞彙表。這樣做是為了加快執行速度，避免讓你等待太久。然而，我們承擔了一些風險，即整個數據集中的某些詞可能不會被包含在詞彙表中，並在訓練過程中被忽略。因此，使用完整的詞彙表大小並在 `adapt` 過程中遍歷整個數據集應該可以提高最終的準確性，但提升幅度不會太大。\n",
|
||||
"\n",
|
||||
"現在我們可以訪問實際的詞彙表：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 6,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"['', '[UNK]', 'the', 'to', 'a', 'in', 'of', 'and', 'on', 'for']\n",
|
||||
"Length of vocabulary: 5335\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"vocab = vectorizer.get_vocabulary()\n",
|
||||
"vocab_size = len(vocab)\n",
|
||||
"print(vocab[:10])\n",
|
||||
"print(f\"Length of vocabulary: {vocab_size}\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"使用向量化器，我們可以輕鬆地將任何文本編碼為一組數字：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<tf.Tensor: shape=(7,), dtype=int64, numpy=array([ 112, 3695, 3, 304, 11, 1041, 1], dtype=int64)>"
|
||||
]
|
||||
},
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"vectorizer('I love to play with my words')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 詞袋（Bag-of-words）文本表示法\n",
|
||||
"\n",
|
||||
"由於文字代表了意義，有時我們可以僅通過查看單個詞語來理解一段文字的意思，而不需要考慮句子中的詞序。例如，在分類新聞時，像 *weather* 和 *snow* 這樣的詞語可能表明是 *天氣預報*，而像 *stocks* 和 *dollar* 則可能屬於 *財經新聞*。\n",
|
||||
"\n",
|
||||
"**詞袋**（BoW）向量表示法是最容易理解的傳統向量表示法。每個詞語都對應一個向量索引，而向量中的元素則表示某個詞語在特定文檔中出現的次數。\n",
|
||||
"\n",
|
||||
" \n",
|
||||
"\n",
|
||||
"> **Note**: 你也可以將 BoW 理解為文本中每個詞語的單熱編碼（one-hot-encoded）向量的總和。\n",
|
||||
"\n",
|
||||
"以下是一個使用 Scikit Learn Python 庫生成詞袋表示法的範例：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 8,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"array([[1, 1, 0, 2, 0, 0, 0, 0, 0]], dtype=int64)"
|
||||
]
|
||||
},
|
||||
"execution_count": 8,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"from sklearn.feature_extraction.text import CountVectorizer\n",
|
||||
"sc_vectorizer = CountVectorizer()\n",
|
||||
"corpus = [\n",
|
||||
" 'I like hot dogs.',\n",
|
||||
" 'The dog ran fast.',\n",
|
||||
" 'Its hot outside.',\n",
|
||||
" ]\n",
|
||||
"sc_vectorizer.fit_transform(corpus)\n",
|
||||
"sc_vectorizer.transform(['My dog likes hot dogs on a hot day.']).toarray()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們也可以使用我們在上面定義的 Keras 向量化器，將每個單詞編號轉換為一個獨熱編碼，然後將所有這些向量相加：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 9,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"array([0., 5., 0., ..., 0., 0., 0.], dtype=float32)"
|
||||
]
|
||||
},
|
||||
"execution_count": 9,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"def to_bow(text):\n",
|
||||
" return tf.reduce_sum(tf.one_hot(vectorizer(text),vocab_size),axis=0)\n",
|
||||
"\n",
|
||||
"to_bow('My dog likes hot dogs on a hot day.').numpy()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"> **注意**：你可能會驚訝地發現結果與之前的例子有所不同。原因是，在 Keras 的例子中，向量的長度對應於詞彙表的大小，而這個詞彙表是基於整個 AG News 數據集建立的；而在 Scikit Learn 的例子中，我們是即時從樣本文本中建立詞彙表的。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 訓練 BoW 分類器\n",
|
||||
"\n",
|
||||
"現在我們已經學會如何建立文字的詞袋表示法，接下來讓我們訓練一個使用該表示法的分類器。首先，我們需要將數據集轉換為詞袋表示法。這可以通過以下方式使用 `map` 函數來實現：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 11,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"batch_size = 128\n",
|
||||
"\n",
|
||||
"ds_train_bow = ds_train.map(lambda x: (to_bow(x['title']+x['description']),x['label'])).batch(batch_size)\n",
|
||||
"ds_test_bow = ds_test.map(lambda x: (to_bow(x['title']+x['description']),x['label'])).batch(batch_size)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在讓我們定義一個簡單的分類器神經網絡，其中包含一個線性層。輸入大小為 `vocab_size`，輸出大小對應於類別數量（4）。由於我們正在解決分類任務，最終的激活函數是 **softmax**：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 12,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"938/938 [==============================] - 66s 70ms/step - loss: 0.6144 - acc: 0.8427 - val_loss: 0.4416 - val_acc: 0.8697\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<keras.callbacks.History at 0x20c70a947f0>"
|
||||
]
|
||||
},
|
||||
"execution_count": 12,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model = keras.models.Sequential([\n",
|
||||
" keras.layers.Dense(4,activation='softmax',input_shape=(vocab_size,))\n",
|
||||
"])\n",
|
||||
"model.compile(loss='sparse_categorical_crossentropy',optimizer='adam',metrics=['acc'])\n",
|
||||
"model.fit(ds_train_bow,validation_data=ds_test_bow)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"由於我們有四個類別，超過 80% 的準確率已經是一個不錯的結果。\n",
|
||||
"\n",
|
||||
"## 將分類器作為一個網絡進行訓練\n",
|
||||
"\n",
|
||||
"由於向量化器也是一個 Keras 層，我們可以定義一個包含它的網絡，並進行端到端的訓練。這樣我們就不需要使用 `map` 來向量化數據集，只需將原始數據集傳遞到網絡的輸入即可。\n",
|
||||
"\n",
|
||||
"> **注意**：我們仍然需要對數據集應用 map 操作，將字典中的字段（例如 `title`、`description` 和 `label`）轉換為元組。然而，當從磁碟加載數據時，我們可以一開始就構建具有所需結構的數據集。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 13,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Model: \"model\"\n",
|
||||
"_________________________________________________________________\n",
|
||||
" Layer (type) Output Shape Param # \n",
|
||||
"=================================================================\n",
|
||||
" input_1 (InputLayer) [(None, 1)] 0 \n",
|
||||
" \n",
|
||||
" text_vectorization (TextVec (None, None) 0 \n",
|
||||
" torization) \n",
|
||||
" \n",
|
||||
" tf.one_hot (TFOpLambda) (None, None, 5335) 0 \n",
|
||||
" \n",
|
||||
" tf.math.reduce_sum (TFOpLam (None, 5335) 0 \n",
|
||||
" bda) \n",
|
||||
" \n",
|
||||
" dense_2 (Dense) (None, 4) 21344 \n",
|
||||
" \n",
|
||||
"=================================================================\n",
|
||||
"Total params: 21,344\n",
|
||||
"Trainable params: 21,344\n",
|
||||
"Non-trainable params: 0\n",
|
||||
"_________________________________________________________________\n",
|
||||
"938/938 [==============================] - 73s 77ms/step - loss: 0.6057 - acc: 0.8414 - val_loss: 0.4202 - val_acc: 0.8736\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<keras.callbacks.History at 0x20c721521f0>"
|
||||
]
|
||||
},
|
||||
"execution_count": 13,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"def extract_text(x):\n",
|
||||
" return x['title']+' '+x['description']\n",
|
||||
"\n",
|
||||
"def tupelize(x):\n",
|
||||
" return (extract_text(x),x['label'])\n",
|
||||
"\n",
|
||||
"inp = keras.Input(shape=(1,),dtype=tf.string)\n",
|
||||
"x = vectorizer(inp)\n",
|
||||
"x = tf.reduce_sum(tf.one_hot(x,vocab_size),axis=1)\n",
|
||||
"out = keras.layers.Dense(4,activation='softmax')(x)\n",
|
||||
"model = keras.models.Model(inp,out)\n",
|
||||
"model.summary()\n",
|
||||
"\n",
|
||||
"model.compile(loss='sparse_categorical_crossentropy',optimizer='adam',metrics=['acc'])\n",
|
||||
"model.fit(ds_train.map(tupelize).batch(batch_size),validation_data=ds_test.map(tupelize).batch(batch_size))\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 雙字組、三字組和 n 字組\n",
|
||||
"\n",
|
||||
"袋裝詞語方法的一個限制是，有些詞語屬於多詞表達，例如「熱狗」這個詞的意思與「熱」和「狗」在其他語境中的意思完全不同。如果我們總是用相同的向量來表示「熱」和「狗」，可能會令模型感到混淆。\n",
|
||||
"\n",
|
||||
"為了解決這個問題，**n 字組表示法**經常用於文件分類的方法中，其中每個詞語、雙詞或三詞的頻率是訓練分類器的一個有用特徵。例如，在雙字組表示法中，我們會將所有詞語配對加入詞彙表，除了原本的詞語之外。\n",
|
||||
"\n",
|
||||
"以下是一個使用 Scikit Learn 生成雙字組袋裝詞語表示法的例子：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 14,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Vocabulary:\n",
|
||||
" {'i': 7, 'like': 11, 'hot': 4, 'dogs': 2, 'i like': 8, 'like hot': 12, 'hot dogs': 5, 'the': 16, 'dog': 0, 'ran': 14, 'fast': 3, 'the dog': 17, 'dog ran': 1, 'ran fast': 15, 'its': 9, 'outside': 13, 'its hot': 10, 'hot outside': 6}\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"array([[1, 0, 1, 0, 2, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]],\n",
|
||||
" dtype=int64)"
|
||||
]
|
||||
},
|
||||
"execution_count": 14,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"bigram_vectorizer = CountVectorizer(ngram_range=(1, 2), token_pattern=r'\\b\\w+\\b', min_df=1)\n",
|
||||
"corpus = [\n",
|
||||
" 'I like hot dogs.',\n",
|
||||
" 'The dog ran fast.',\n",
|
||||
" 'Its hot outside.',\n",
|
||||
" ]\n",
|
||||
"bigram_vectorizer.fit_transform(corpus)\n",
|
||||
"print(\"Vocabulary:\\n\",bigram_vectorizer.vocabulary_)\n",
|
||||
"bigram_vectorizer.transform(['My dog likes hot dogs on a hot day.']).toarray()\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"n-gram 方法的主要缺點是詞彙量會迅速膨脹。在實際應用中，我們需要將 n-gram 表示與降維技術結合使用，例如 *嵌入*，我們會在下一單元中討論。\n",
|
||||
"\n",
|
||||
"要在我們的 **AG News** 數據集中使用 n-gram 表示，我們需要將 `ngrams` 參數傳遞給 `TextVectorization` 建構函數。二元語法詞彙的長度**顯著更大**，在我們的情況下，超過 130 萬個詞元！因此，限制二元語法詞元的數量至合理範圍是有道理的。\n",
|
||||
"\n",
|
||||
"我們可以使用與上面相同的程式碼來訓練分類器，但這樣會非常耗費記憶體。在下一單元中，我們將使用嵌入來訓練二元語法分類器。同時，你可以在這個筆記本中嘗試訓練二元語法分類器，看看是否能獲得更高的準確率。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 自動計算 BoW 向量\n",
|
||||
"\n",
|
||||
"在上述例子中，我們透過手動方式計算 BoW 向量，方法是將個別詞語的一次性編碼相加。然而，最新版本的 TensorFlow 允許我們透過在向量化器構造函數中傳入 `output_mode='count` 參數，自動計算 BoW 向量。這使得定義和訓練模型變得更加簡單：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 15,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Training vectorizer\n",
|
||||
"938/938 [==============================] - 7s 7ms/step - loss: 0.5929 - acc: 0.8486 - val_loss: 0.4168 - val_acc: 0.8772\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<keras.callbacks.History at 0x20c725217c0>"
|
||||
]
|
||||
},
|
||||
"execution_count": 15,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model = keras.models.Sequential([\n",
|
||||
" keras.layers.experimental.preprocessing.TextVectorization(max_tokens=vocab_size,output_mode='count'),\n",
|
||||
" keras.layers.Dense(4,input_shape=(vocab_size,), activation='softmax')\n",
|
||||
"])\n",
|
||||
"print(\"Training vectorizer\")\n",
|
||||
"model.layers[0].adapt(ds_train.take(500).map(extract_text))\n",
|
||||
"model.compile(loss='sparse_categorical_crossentropy',optimizer='adam',metrics=['acc'])\n",
|
||||
"model.fit(ds_train.map(tupelize).batch(batch_size),validation_data=ds_test.map(tupelize).batch(batch_size))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 詞頻 - 逆文件頻率 (TF-IDF)\n",
|
||||
"\n",
|
||||
"在詞袋（BoW）表示法中，詞語的出現次數是使用相同的技術進行加權，而不考慮詞語本身。然而，很明顯像 *a* 和 *in* 這些常見詞語對分類的作用遠不如專業術語。在大多數自然語言處理（NLP）任務中，有些詞語比其他詞語更重要。\n",
|
||||
"\n",
|
||||
"**TF-IDF** 代表 **詞頻 - 逆文件頻率**。這是一種詞袋的變體，其中不是用二進制的 0/1 值來表示詞語是否出現在文件中，而是使用浮點值，該值與詞語在語料庫中的出現頻率相關。\n",
|
||||
"\n",
|
||||
"更正式地說，詞語 $i$ 在文件 $j$ 中的權重 $w_{ij}$ 定義為：\n",
|
||||
"$$\n",
|
||||
"w_{ij} = tf_{ij}\\times\\log({N\\over df_i})\n",
|
||||
"$$\n",
|
||||
"其中：\n",
|
||||
"* $tf_{ij}$ 是詞語 $i$ 在文件 $j$ 中的出現次數，即我們之前看到的詞袋值\n",
|
||||
"* $N$ 是集合中的文件數量\n",
|
||||
"* $df_i$ 是整個集合中包含詞語 $i$ 的文件數量\n",
|
||||
"\n",
|
||||
"TF-IDF 值 $w_{ij}$ 隨著詞語在文件中出現次數的增加而增加，同時會因語料庫中包含該詞語的文件數量而進行調整，這有助於平衡某些詞語出現頻率較高的情況。例如，如果某個詞語出現在集合中的 *每一個* 文件中，則 $df_i=N$，而 $w_{ij}=0$，這些詞語將被完全忽略。\n",
|
||||
"\n",
|
||||
"你可以使用 Scikit Learn 輕鬆地創建文本的 TF-IDF 向量化：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 16,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"array([[0.43381609, 0. , 0.43381609, 0. , 0.65985664,\n",
|
||||
" 0.43381609, 0. , 0. , 0. , 0. ,\n",
|
||||
" 0. , 0. , 0. , 0. , 0. ,\n",
|
||||
" 0. ]])"
|
||||
]
|
||||
},
|
||||
"execution_count": 16,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"from sklearn.feature_extraction.text import TfidfVectorizer\n",
|
||||
"vectorizer = TfidfVectorizer(ngram_range=(1,2))\n",
|
||||
"vectorizer.fit_transform(corpus)\n",
|
||||
"vectorizer.transform(['My dog likes hot dogs on a hot day.']).toarray()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"在 Keras 中，`TextVectorization` 層可以通過傳遞 `output_mode='tf-idf'` 參數自動計算 TF-IDF 頻率。我們重複上面使用的代碼，看看使用 TF-IDF 是否能提高準確性：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 17,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Training vectorizer\n",
|
||||
"938/938 [==============================] - 12s 12ms/step - loss: 0.4197 - acc: 0.8662 - val_loss: 0.3432 - val_acc: 0.8849\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<keras.callbacks.History at 0x20c729dfd30>"
|
||||
]
|
||||
},
|
||||
"execution_count": 17,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model = keras.models.Sequential([\n",
|
||||
" keras.layers.experimental.preprocessing.TextVectorization(max_tokens=vocab_size,output_mode='tf-idf'),\n",
|
||||
" keras.layers.Dense(4,input_shape=(vocab_size,), activation='softmax')\n",
|
||||
"])\n",
|
||||
"print(\"Training vectorizer\")\n",
|
||||
"model.layers[0].adapt(ds_train.take(500).map(extract_text))\n",
|
||||
"model.compile(loss='sparse_categorical_crossentropy',optimizer='adam',metrics=['acc'])\n",
|
||||
"model.fit(ds_train.map(tupelize).batch(batch_size),validation_data=ds_test.map(tupelize).batch(batch_size))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 結論\n",
|
||||
"\n",
|
||||
"雖然 TF-IDF 表示法為不同的詞語提供了頻率權重，但它無法表達詞語的意義或順序。正如著名語言學家 J. R. Firth 在1935年所說：「詞語的完整意義總是與上下文相關，任何脫離上下文的意義研究都不應被認真對待。」我們稍後在課程中將學習如何通過語言建模來捕捉文本中的上下文信息。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件已使用人工智能翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。雖然我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。原始語言的文件應被視為權威來源。對於重要資訊，建議使用專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或錯誤解釋概不負責。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"interpreter": {
|
||||
"hash": "0cb620c6d4b9f7a635928804c26cf22403d89d98d79684e4529119355ee6d5a5"
|
||||
},
|
||||
"kernel_info": {
|
||||
"name": "conda-env-py37_tensorflow-py"
|
||||
},
|
||||
"kernelspec": {
|
||||
"display_name": "py37_tensorflow",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.8.12"
|
||||
},
|
||||
"nteract": {
|
||||
"version": "nteract-front-end@1.0.0"
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "19b43951d55b377a76209c24c1f017e4",
|
||||
"translation_date": "2025-08-31T11:01:35+00:00",
|
||||
"source_file": "lessons/5-NLP/13-TextRep/TextRepresentationTF.ipynb",
|
||||
"language_code": "hk"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 4
|
||||
}
|
||||
|
|
@ -0,0 +1,724 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 嵌入\n",
|
||||
"\n",
|
||||
"在之前的例子中，我們使用了高維度的詞袋向量，其長度為 `vocab_size`，並且我們明確地將低維度的位置信息向量轉換為稀疏的獨熱表示。這種獨熱表示並不具備記憶效率，此外，每個詞都被獨立處理，也就是說，獨熱編碼的向量無法表達詞與詞之間的語義相似性。\n",
|
||||
"\n",
|
||||
"在這一單元中，我們將繼續探索 **News AG** 數據集。首先，讓我們載入數據並從之前的筆記本中獲取一些定義。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 1,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Loading dataset...\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"name": "stderr",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"d:\\WORK\\ai-for-beginners\\5-NLP\\14-Embeddings\\data\\train.csv: 29.5MB [00:01, 18.8MB/s] \n",
|
||||
"d:\\WORK\\ai-for-beginners\\5-NLP\\14-Embeddings\\data\\test.csv: 1.86MB [00:00, 11.2MB/s] \n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Building vocab...\n",
|
||||
"Vocab size = 95812\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"import torch\n",
|
||||
"import torchtext\n",
|
||||
"import numpy as np\n",
|
||||
"from torchnlp import *\n",
|
||||
"train_dataset, test_dataset, classes, vocab = load_dataset()\n",
|
||||
"vocab_size = len(vocab)\n",
|
||||
"print(\"Vocab size = \",vocab_size)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 什麼是嵌入？\n",
|
||||
"\n",
|
||||
"**嵌入**的概念是用低維度的密集向量來表示單詞，這些向量能夠在某種程度上反映單詞的語義。我們稍後會討論如何構建有意義的單詞嵌入，但現在我們可以簡單地將嵌入理解為一種降低單詞向量維度的方法。\n",
|
||||
"\n",
|
||||
"因此，嵌入層會將一個單詞作為輸入，並生成一個指定 `embedding_size` 的輸出向量。從某種意義上說，它與 `Linear` 層非常相似，但不同的是，它不是接收 one-hot 編碼的向量，而是能夠接收一個單詞的編號作為輸入。\n",
|
||||
"\n",
|
||||
"通過將嵌入層作為我們網絡的第一層，我們可以從詞袋模型（bag-of-words）切換到 **嵌入袋模型**（embedding bag model）。在這個模型中，我們首先將文本中的每個單詞轉換為對應的嵌入，然後對所有這些嵌入執行某種聚合函數，例如 `sum`、`average` 或 `max`。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"我們的分類器神經網絡將以嵌入層開始，接著是聚合層，最後在其上添加一個線性分類器：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 2,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"class EmbedClassifier(torch.nn.Module):\n",
|
||||
" def __init__(self, vocab_size, embed_dim, num_class):\n",
|
||||
" super().__init__()\n",
|
||||
" self.embedding = torch.nn.Embedding(vocab_size, embed_dim)\n",
|
||||
" self.fc = torch.nn.Linear(embed_dim, num_class)\n",
|
||||
"\n",
|
||||
" def forward(self, x):\n",
|
||||
" x = self.embedding(x)\n",
|
||||
" x = torch.mean(x,dim=1)\n",
|
||||
" return self.fc(x)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"### 處理變化的序列大小\n",
|
||||
"\n",
|
||||
"由於這種架構，我們需要以特定方式建立送入網絡的迷你批次。在上一單元中，使用詞袋模型（bag-of-words）時，迷你批次中的所有 BoW 張量都具有相同的大小 `vocab_size`，無論文本序列的實際長度是多少。一旦我們轉向使用詞嵌入（word embeddings），每個文本樣本中的詞數會有所不同，而在將這些樣本合併成迷你批次時，我們需要進行一些填充。\n",
|
||||
"\n",
|
||||
"這可以通過向數據源提供 `collate_fn` 函數來完成：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"def padify(b):\n",
|
||||
" # b is the list of tuples of length batch_size\n",
|
||||
" # - first element of a tuple = label, \n",
|
||||
" # - second = feature (text sequence)\n",
|
||||
" # build vectorized sequence\n",
|
||||
" v = [encode(x[1]) for x in b]\n",
|
||||
" # first, compute max length of a sequence in this minibatch\n",
|
||||
" l = max(map(len,v))\n",
|
||||
" return ( # tuple of two tensors - labels and features\n",
|
||||
" torch.LongTensor([t[0]-1 for t in b]),\n",
|
||||
" torch.stack([torch.nn.functional.pad(torch.tensor(t),(0,l-len(t)),mode='constant',value=0) for t in v])\n",
|
||||
" )\n",
|
||||
"\n",
|
||||
"train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=16, collate_fn=padify, shuffle=True)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"### 訓練嵌入分類器\n",
|
||||
"\n",
|
||||
"現在我們已經定義了合適的數據加載器，我們可以使用上一單元中定義的訓練函數來訓練模型：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"3200: acc=0.6415625\n",
|
||||
"6400: acc=0.6865625\n",
|
||||
"9600: acc=0.7103125\n",
|
||||
"12800: acc=0.726953125\n",
|
||||
"16000: acc=0.739375\n",
|
||||
"19200: acc=0.75046875\n",
|
||||
"22400: acc=0.7572321428571429\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"(0.889799795315499, 0.7623160588611644)"
|
||||
]
|
||||
},
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"net = EmbedClassifier(vocab_size,32,len(classes)).to(device)\n",
|
||||
"train_epoch(net,train_loader, lr=1, epoch_size=25000)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"> **注意**：我們在這裡僅訓練 25k 條記錄（少於一個完整的 epoch）以節省時間，但您可以繼續訓練，編寫一個函數來訓練多個 epoch，並嘗試調整學習率參數以獲得更高的準確率。您應該能夠達到約 90% 的準確率。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"### EmbeddingBag 層與可變長度序列表示法\n",
|
||||
"\n",
|
||||
"在之前的架構中，我們需要將所有序列填充至相同的長度，以便將它們放入小批量中。這並不是表示可變長度序列的最有效方法——另一種方法是使用 **offset** 向量，該向量會保存所有序列在一個大型向量中的偏移量。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"> **Note**: 在上圖中，我們展示的是字符序列，但在我們的例子中，我們處理的是單詞序列。然而，用偏移向量表示序列的基本原則仍然相同。\n",
|
||||
"\n",
|
||||
"為了使用偏移表示法，我們使用 [`EmbeddingBag`](https://pytorch.org/docs/stable/generated/torch.nn.EmbeddingBag.html) 層。它類似於 `Embedding`，但它以內容向量和偏移向量作為輸入，並且還包含一個平均層，可以是 `mean`、`sum` 或 `max`。\n",
|
||||
"\n",
|
||||
"以下是使用 `EmbeddingBag` 的修改後的網絡：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"class EmbedClassifier(torch.nn.Module):\n",
|
||||
" def __init__(self, vocab_size, embed_dim, num_class):\n",
|
||||
" super().__init__()\n",
|
||||
" self.embedding = torch.nn.EmbeddingBag(vocab_size, embed_dim)\n",
|
||||
" self.fc = torch.nn.Linear(embed_dim, num_class)\n",
|
||||
"\n",
|
||||
" def forward(self, text, off):\n",
|
||||
" x = self.embedding(text, off)\n",
|
||||
" return self.fc(x)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"要準備數據集進行訓練，我們需要提供一個轉換函數來準備偏移向量：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 6,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"def offsetify(b):\n",
|
||||
" # first, compute data tensor from all sequences\n",
|
||||
" x = [torch.tensor(encode(t[1])) for t in b]\n",
|
||||
" # now, compute the offsets by accumulating the tensor of sequence lengths\n",
|
||||
" o = [0] + [len(t) for t in x]\n",
|
||||
" o = torch.tensor(o[:-1]).cumsum(dim=0)\n",
|
||||
" return ( \n",
|
||||
" torch.LongTensor([t[0]-1 for t in b]), # labels\n",
|
||||
" torch.cat(x), # text \n",
|
||||
" o\n",
|
||||
" )\n",
|
||||
"\n",
|
||||
"train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=16, collate_fn=offsetify, shuffle=True)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"注意，與之前的所有例子不同，我們的網絡現在接受兩個參數：數據向量和偏移向量，這兩者的大小不同。同樣，我們的數據加載器現在也提供了3個值而不是2個：文本和偏移向量都作為特徵提供。因此，我們需要稍微調整我們的訓練函數來處理這一點：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"3200: acc=0.6153125\n",
|
||||
"6400: acc=0.6615625\n",
|
||||
"9600: acc=0.6932291666666667\n",
|
||||
"12800: acc=0.715078125\n",
|
||||
"16000: acc=0.7270625\n",
|
||||
"19200: acc=0.7382291666666667\n",
|
||||
"22400: acc=0.7486160714285715\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"(22.771553103007037, 0.7551983365323096)"
|
||||
]
|
||||
},
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"net = EmbedClassifier(vocab_size,32,len(classes)).to(device)\n",
|
||||
"\n",
|
||||
"def train_epoch_emb(net,dataloader,lr=0.01,optimizer=None,loss_fn = torch.nn.CrossEntropyLoss(),epoch_size=None, report_freq=200):\n",
|
||||
" optimizer = optimizer or torch.optim.Adam(net.parameters(),lr=lr)\n",
|
||||
" loss_fn = loss_fn.to(device)\n",
|
||||
" net.train()\n",
|
||||
" total_loss,acc,count,i = 0,0,0,0\n",
|
||||
" for labels,text,off in dataloader:\n",
|
||||
" optimizer.zero_grad()\n",
|
||||
" labels,text,off = labels.to(device), text.to(device), off.to(device)\n",
|
||||
" out = net(text, off)\n",
|
||||
" loss = loss_fn(out,labels) #cross_entropy(out,labels)\n",
|
||||
" loss.backward()\n",
|
||||
" optimizer.step()\n",
|
||||
" total_loss+=loss\n",
|
||||
" _,predicted = torch.max(out,1)\n",
|
||||
" acc+=(predicted==labels).sum()\n",
|
||||
" count+=len(labels)\n",
|
||||
" i+=1\n",
|
||||
" if i%report_freq==0:\n",
|
||||
" print(f\"{count}: acc={acc.item()/count}\")\n",
|
||||
" if epoch_size and count>epoch_size:\n",
|
||||
" break\n",
|
||||
" return total_loss.item()/count, acc.item()/count\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"train_epoch_emb(net,train_loader, lr=4, epoch_size=25000)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 語意嵌入：Word2Vec\n",
|
||||
"\n",
|
||||
"在我們之前的例子中，模型的嵌入層學會了將單詞映射到向量表示，但這種表示並沒有太多語意上的意義。如果能學習到這樣的向量表示，讓相似的單詞或同義詞在某種向量距離（例如歐幾里得距離）上彼此接近，那就更理想了。\n",
|
||||
"\n",
|
||||
"為了實現這一點，我們需要以特定的方式在大量文本數據上預訓練嵌入模型。最早的語意嵌入訓練方法之一被稱為 [Word2Vec](https://en.wikipedia.org/wiki/Word2vec)。它基於兩種主要架構，用於生成單詞的分佈式表示：\n",
|
||||
"\n",
|
||||
" - **連續詞袋模型** (CBoW) — 在這種架構中，我們訓練模型根據周圍的上下文來預測一個單詞。給定 ngram $(W_{-2},W_{-1},W_0,W_1,W_2)$，模型的目標是從 $(W_{-2},W_{-1},W_1,W_2)$ 預測 $W_0$。\n",
|
||||
" - **連續跳字模型** (Skip-Gram) 與 CBoW 相反。模型使用周圍的上下文窗口單詞來預測當前單詞。\n",
|
||||
"\n",
|
||||
"CBoW 的速度較快，而 Skip-Gram 雖然較慢，但在表示不常見單詞方面表現更好。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"為了試驗在 Google News 數據集上預訓練的 Word2Vec 嵌入，我們可以使用 **gensim** 庫。以下是找到與 'neural' 最相似的單詞的例子：\n",
|
||||
"\n",
|
||||
"> **注意：** 當你第一次創建單詞向量時，下載它們可能需要一些時間！\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 8,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import gensim.downloader as api\n",
|
||||
"w2v = api.load('word2vec-google-news-300')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 9,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"neuronal -> 0.7804799675941467\n",
|
||||
"neurons -> 0.7326500415802002\n",
|
||||
"neural_circuits -> 0.7252851724624634\n",
|
||||
"neuron -> 0.7174385190010071\n",
|
||||
"cortical -> 0.6941086649894714\n",
|
||||
"brain_circuitry -> 0.6923246383666992\n",
|
||||
"synaptic -> 0.6699118614196777\n",
|
||||
"neural_circuitry -> 0.6638563275337219\n",
|
||||
"neurochemical -> 0.6555314064025879\n",
|
||||
"neuronal_activity -> 0.6531826257705688\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"for w,p in w2v.most_similar('neural'):\n",
|
||||
" print(f\"{w} -> {p}\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們亦可從該詞計算向量嵌入，用於訓練分類模型（為清晰起見，我們僅顯示向量的前20個組件）：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 10,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"array([ 0.01226807, 0.06225586, 0.10693359, 0.05810547, 0.23828125,\n",
|
||||
" 0.03686523, 0.05151367, -0.20703125, 0.01989746, 0.10058594,\n",
|
||||
" -0.03759766, -0.1015625 , -0.15820312, -0.08105469, -0.0390625 ,\n",
|
||||
" -0.05053711, 0.16015625, 0.2578125 , 0.10058594, -0.25976562],\n",
|
||||
" dtype=float32)"
|
||||
]
|
||||
},
|
||||
"execution_count": 10,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"w2v.word_vec('play')[:20]"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"語義嵌入的優點在於你可以操作向量編碼來改變語義。例如，我們可以要求找到一個詞，其向量表示盡可能接近詞語*國王*和*女人*，並且盡可能遠離詞語*男人*：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 10,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"('queen', 0.7118192911148071)"
|
||||
]
|
||||
},
|
||||
"execution_count": 10,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"w2v.most_similar(positive=['king','woman'],negative=['man'])[0]"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"CBoW 和 Skip-Grams 都是「預測型」的詞嵌入方法，因為它們只考慮局部上下文。Word2Vec 並沒有利用全局上下文。\n",
|
||||
"\n",
|
||||
"**FastText** 在 Word2Vec 的基礎上進一步發展，通過學習每個詞的向量表示以及詞內的字符 n-grams。這些表示的值在每次訓練步驟中會被平均成一個向量。雖然這增加了預訓練的計算量，但它使詞嵌入能夠編碼子詞資訊。\n",
|
||||
"\n",
|
||||
"另一種方法，**GloVe**，利用了共現矩陣的概念，通過神經方法將共現矩陣分解為更具表達力和非線性的詞向量。\n",
|
||||
"\n",
|
||||
"你可以通過將嵌入模型切換為 FastText 和 GloVe 來試驗這些例子，因為 gensim 支援多種不同的詞嵌入模型。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 在 PyTorch 中使用預訓練的嵌入\n",
|
||||
"\n",
|
||||
"我們可以修改上述範例，將嵌入層中的矩陣預先填入語義嵌入，例如 Word2Vec。我們需要考慮到，預訓練嵌入的詞彙表與我們文本語料庫的詞彙表可能不一致，因此我們會用隨機值初始化缺失詞彙的權重：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 11,
|
||||
"metadata": {
|
||||
"tags": []
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Embedding size: 300\n",
|
||||
"Populating matrix, this will take some time...Done, found 41080 words, 54732 words missing\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"embed_size = len(w2v.get_vector('hello'))\n",
|
||||
"print(f'Embedding size: {embed_size}')\n",
|
||||
"\n",
|
||||
"net = EmbedClassifier(vocab_size,embed_size,len(classes))\n",
|
||||
"\n",
|
||||
"print('Populating matrix, this will take some time...',end='')\n",
|
||||
"found, not_found = 0,0\n",
|
||||
"for i,w in enumerate(vocab.get_itos()):\n",
|
||||
" try:\n",
|
||||
" net.embedding.weight[i].data = torch.tensor(w2v.get_vector(w))\n",
|
||||
" found+=1\n",
|
||||
" except:\n",
|
||||
" net.embedding.weight[i].data = torch.normal(0.0,1.0,(embed_size,))\n",
|
||||
" not_found+=1\n",
|
||||
"\n",
|
||||
"print(f\"Done, found {found} words, {not_found} words missing\")\n",
|
||||
"net = net.to(device)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在讓我們訓練模型。請注意，由於嵌入層的大小更大，因此參數的數量也大幅增加，訓練模型所需的時間比之前的例子顯著增加。此外，正因如此，如果我們想避免過度擬合，可能需要在更多的例子上訓練模型。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 12,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"3200: acc=0.6359375\n",
|
||||
"6400: acc=0.68109375\n",
|
||||
"9600: acc=0.7067708333333333\n",
|
||||
"12800: acc=0.723671875\n",
|
||||
"16000: acc=0.73625\n",
|
||||
"19200: acc=0.7463541666666667\n",
|
||||
"22400: acc=0.7560714285714286\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"(214.1013875559821, 0.7626759436980166)"
|
||||
]
|
||||
},
|
||||
"execution_count": 12,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"train_epoch_emb(net,train_loader, lr=4, epoch_size=25000)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"在我們的情況中，準確度並沒有顯著提升，這可能是由於詞彙差異較大所致。 \n",
|
||||
"為了解決詞彙差異的問題，我們可以採用以下其中一種解決方案： \n",
|
||||
"* 重新訓練 word2vec 模型，使用我們的詞彙 \n",
|
||||
"* 使用預訓練的 word2vec 模型的詞彙來載入我們的數據集。在載入數據集時，可以指定使用的詞彙。 \n",
|
||||
"\n",
|
||||
"後者的方法似乎更簡單，特別是因為 PyTorch 的 `torchtext` 框架內建了對嵌入的支持。例如，我們可以以下列方式實例化基於 GloVe 的詞彙： \n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 14,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stderr",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"100%|█████████▉| 399999/400000 [00:15<00:00, 25411.14it/s]\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"vocab = torchtext.vocab.GloVe(name='6B', dim=50)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"已載入的詞彙具有以下基本操作：\n",
|
||||
"* `vocab.stoi` 字典可讓我們將單詞轉換為其字典索引\n",
|
||||
"* `vocab.itos` 則相反 - 將數字轉換為單詞\n",
|
||||
"* `vocab.vectors` 是嵌入向量的陣列，因此要獲取單詞 `s` 的嵌入，我們需要使用 `vocab.vectors[vocab.stoi[s]]`\n",
|
||||
"\n",
|
||||
"以下是一個操作嵌入的例子，用以展示方程式 **kind-man+woman = queen**（我稍微調整了一下係數以使其生效）：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 15,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"'queen'"
|
||||
]
|
||||
},
|
||||
"execution_count": 15,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"# get the vector corresponding to kind-man+woman\n",
|
||||
"qvec = vocab.vectors[vocab.stoi['king']]-vocab.vectors[vocab.stoi['man']]+1.3*vocab.vectors[vocab.stoi['woman']]\n",
|
||||
"# find the index of the closest embedding vector \n",
|
||||
"d = torch.sum((vocab.vectors-qvec)**2,dim=1)\n",
|
||||
"min_idx = torch.argmin(d)\n",
|
||||
"# find the corresponding word\n",
|
||||
"vocab.itos[min_idx]"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"要使用這些嵌入來訓練分類器，我們首先需要使用GloVe詞彙表對我們的數據集進行編碼：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 16,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"def offsetify(b):\n",
|
||||
" # first, compute data tensor from all sequences\n",
|
||||
" x = [torch.tensor(encode(t[1],voc=vocab)) for t in b] # pass the instance of vocab to encode function!\n",
|
||||
" # now, compute the offsets by accumulating the tensor of sequence lengths\n",
|
||||
" o = [0] + [len(t) for t in x]\n",
|
||||
" o = torch.tensor(o[:-1]).cumsum(dim=0)\n",
|
||||
" return ( \n",
|
||||
" torch.LongTensor([t[0]-1 for t in b]), # labels\n",
|
||||
" torch.cat(x), # text \n",
|
||||
" o\n",
|
||||
" )"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"正如我們上面所見，所有向量嵌入都存儲在 `vocab.vectors` 矩陣中。這使得通過簡單的複製將這些權重加載到嵌入層的權重中變得非常容易：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 17,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"net = EmbedClassifier(len(vocab),len(vocab.vectors[0]),len(classes))\n",
|
||||
"net.embedding.weight.data = vocab.vectors\n",
|
||||
"net = net.to(device)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": []
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 18,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"3200: acc=0.6271875\n",
|
||||
"6400: acc=0.68078125\n",
|
||||
"9600: acc=0.7030208333333333\n",
|
||||
"12800: acc=0.71984375\n",
|
||||
"16000: acc=0.7346875\n",
|
||||
"19200: acc=0.7455729166666667\n",
|
||||
"22400: acc=0.7529464285714286\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"(35.53972978646833, 0.7575175943698017)"
|
||||
]
|
||||
},
|
||||
"execution_count": 18,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=16, collate_fn=offsetify, shuffle=True)\n",
|
||||
"train_epoch_emb(net,train_loader, lr=4, epoch_size=25000)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們未能看到準確性顯著提高的其中一個原因是因為我們的數據集中的某些詞語在預訓練的GloVe詞彙表中缺失，因此它們基本上被忽略了。為了克服這一問題，我們可以在我們的數據集上訓練自己的嵌入。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 語境嵌入\n",
|
||||
"\n",
|
||||
"傳統預訓練嵌入表示（例如 Word2Vec）的一個主要限制是詞義消歧的問題。雖然預訓練嵌入可以捕捉到一些詞語在語境中的含義，但每個詞的所有可能含義都被編碼到同一個嵌入中。這可能會在下游模型中引發問題，因為許多詞語（例如 \"play\"）的含義會根據使用的語境而有所不同。\n",
|
||||
"\n",
|
||||
"例如，\"play\" 在以下兩個句子中的含義就完全不同：\n",
|
||||
"- 我去劇院看了一場**戲劇**。\n",
|
||||
"- 約翰想和他的朋友**玩耍**。\n",
|
||||
"\n",
|
||||
"上述的預訓練嵌入將 \"play\" 的這兩種含義表示為同一個嵌入。為了解決這個限制，我們需要基於**語言模型**來構建嵌入，語言模型是在大量文本語料庫上訓練的，並且*了解*詞語如何在不同語境中組合使用。討論語境嵌入超出了本教程的範圍，但我們會在下一單元討論語言模型時回到這個主題。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件已使用人工智能翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。應以原文文件作為權威來源。對於關鍵資訊，建議尋求專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或誤釋不承擔責任。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"interpreter": {
|
||||
"hash": "0cb620c6d4b9f7a635928804c26cf22403d89d98d79684e4529119355ee6d5a5"
|
||||
},
|
||||
"kernelspec": {
|
||||
"display_name": "py37_pytorch",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.8.12"
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "f50b026abce5cf36783a560ea72cb9b1",
|
||||
"translation_date": "2025-08-31T10:56:44+00:00",
|
||||
"source_file": "lessons/5-NLP/14-Embeddings/EmbeddingsPyTorch.ipynb",
|
||||
"language_code": "hk"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 2
|
||||
}
|
||||
|
|
@ -0,0 +1,695 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 嵌入\n",
|
||||
"\n",
|
||||
"在我們之前的例子中，我們操作的是長度為 `vocab_size` 的高維度詞袋向量，並且我們將低維度的位置表示向量顯式地轉換為稀疏的獨熱表示（one-hot representation）。這種獨熱表示並不具備記憶效率。此外，每個詞彙都被獨立處理，因此獨熱編碼向量無法表達詞彙之間的語義相似性。\n",
|
||||
"\n",
|
||||
"在本單元中，我們將繼續探索 **News AG** 數據集。首先，讓我們載入數據並從上一單元中獲取一些定義。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 2,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import tensorflow as tf\n",
|
||||
"from tensorflow import keras\n",
|
||||
"import tensorflow_datasets as tfds\n",
|
||||
"import numpy as np\n",
|
||||
"\n",
|
||||
"ds_train, ds_test = tfds.load('ag_news_subset').values()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"### 什麼是嵌入？\n",
|
||||
"\n",
|
||||
"**嵌入**的概念是用低維度的密集向量來表示詞語，這些向量反映了詞語的語義。稍後我們會討論如何構建有意義的詞嵌入，但現在可以簡單地將嵌入理解為一種降低詞向量維度的方法。\n",
|
||||
"\n",
|
||||
"嵌入層以詞語作為輸入，並生成指定的 `embedding_size` 的輸出向量。某種程度上，它與 `Dense` 層非常相似，但嵌入層不需要以 one-hot 編碼向量作為輸入，而是可以直接接受詞語編號。\n",
|
||||
"\n",
|
||||
"通過在網絡中使用嵌入層作為第一層，我們可以從詞袋模型切換到 **嵌入袋** 模型。在嵌入袋模型中，我們首先將文本中的每個詞轉換為相應的嵌入，然後對所有嵌入進行某種聚合函數計算，例如 `sum`、`average` 或 `max`。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"我們的分類器神經網絡包含以下幾層：\n",
|
||||
"\n",
|
||||
"* `TextVectorization` 層：該層以字符串作為輸入，並生成包含詞語編號的張量。我們會指定一個合理的詞彙表大小 `vocab_size`，並忽略使用頻率較低的詞語。輸入形狀為 1，輸出形狀為 $n$，因為結果中會有 $n$ 個標記，每個標記包含從 0 到 `vocab_size` 的數字。\n",
|
||||
"* `Embedding` 層：該層接收 $n$ 個數字，並將每個數字縮減為指定長度的密集向量（在我們的例子中是 100）。因此，形狀為 $n$ 的輸入張量將被轉換為 $n\\times 100$ 的張量。\n",
|
||||
"* 聚合層：該層沿第一軸計算該張量的平均值，即計算所有 $n$ 個輸入張量（對應不同詞語）的平均值。為了實現該層，我們會使用 `Lambda` 層，並傳入計算平均值的函數。輸出形狀為 100，這將是整個輸入序列的數字表示。\n",
|
||||
"* 最後的 `Dense` 線性分類器。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Model: \"sequential\"\n",
|
||||
"_________________________________________________________________\n",
|
||||
" Layer (type) Output Shape Param # \n",
|
||||
"=================================================================\n",
|
||||
" text_vectorization (TextVec (None, None) 0 \n",
|
||||
" torization) \n",
|
||||
" \n",
|
||||
" embedding (Embedding) (None, None, 100) 3000000 \n",
|
||||
" \n",
|
||||
" lambda (Lambda) (None, 100) 0 \n",
|
||||
" \n",
|
||||
" dense (Dense) (None, 4) 404 \n",
|
||||
" \n",
|
||||
"=================================================================\n",
|
||||
"Total params: 3,000,404\n",
|
||||
"Trainable params: 3,000,404\n",
|
||||
"Non-trainable params: 0\n",
|
||||
"_________________________________________________________________\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"vocab_size = 30000\n",
|
||||
"batch_size = 128\n",
|
||||
"\n",
|
||||
"vectorizer = keras.layers.experimental.preprocessing.TextVectorization(max_tokens=vocab_size,input_shape=(1,))\n",
|
||||
"\n",
|
||||
"model = keras.models.Sequential([\n",
|
||||
" vectorizer, \n",
|
||||
" keras.layers.Embedding(vocab_size,100),\n",
|
||||
" keras.layers.Lambda(lambda x: tf.reduce_mean(x,axis=1)),\n",
|
||||
" keras.layers.Dense(4, activation='softmax')\n",
|
||||
"])\n",
|
||||
"model.summary()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"在 `summary` 的輸出中，**output shape** 欄位中的第一個張量維度 `None` 代表小批量的大小，第二個維度則代表標記序列的長度。小批量中的所有標記序列長度都不相同。我們會在下一節討論如何處理這個問題。\n",
|
||||
"\n",
|
||||
"現在讓我們開始訓練網絡：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Training vectorizer\n",
|
||||
"938/938 [==============================] - 20s 20ms/step - loss: 0.7891 - acc: 0.8155 - val_loss: 0.4470 - val_acc: 0.8642\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<keras.callbacks.History at 0x22255515100>"
|
||||
]
|
||||
},
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"def extract_text(x):\n",
|
||||
" return x['title']+' '+x['description']\n",
|
||||
"\n",
|
||||
"def tupelize(x):\n",
|
||||
" return (extract_text(x),x['label'])\n",
|
||||
"\n",
|
||||
"print(\"Training vectorizer\")\n",
|
||||
"vectorizer.adapt(ds_train.take(500).map(extract_text))\n",
|
||||
"\n",
|
||||
"model.compile(loss='sparse_categorical_crossentropy',metrics=['acc'])\n",
|
||||
"model.fit(ds_train.map(tupelize).batch(batch_size),validation_data=ds_test.map(tupelize).batch(batch_size))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"nteract": {
|
||||
"transient": {
|
||||
"deleting": false
|
||||
}
|
||||
}
|
||||
},
|
||||
"source": [
|
||||
"> **注意** 我們正在基於數據的一個子集構建向量化器。這樣做是為了加快過程，但可能會導致我們文本中的某些詞彙未出現在詞彙表中。在這種情況下，這些詞彙將被忽略，這可能會導致準確性略有下降。然而，在現實生活中，文本的子集通常能提供良好的詞彙估計。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"### 處理變量序列長度\n",
|
||||
"\n",
|
||||
"讓我們了解小批量訓練的過程。在上述例子中，輸入張量的維度是1，我們使用長度為128的小批量，因此張量的實際大小是 $128 \\times 1$。然而，每個句子的詞元數量是不同的。如果我們將 `TextVectorization` 層應用到單一輸入，返回的詞元數量會根據文本如何被分詞而有所不同：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"tf.Tensor([ 1 45], shape=(2,), dtype=int64)\n",
|
||||
"tf.Tensor([ 112 1271 1 3 1747 158], shape=(6,), dtype=int64)\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"print(vectorizer('Hello, world!'))\n",
|
||||
"print(vectorizer('I am glad to meet you!'))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"然而，當我們將向量化器應用於多個序列時，它必須生成一個矩形形狀的張量，因此會用 PAD 標記（在我們的情況下是零）填充未使用的元素：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 6,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<tf.Tensor: shape=(2, 6), dtype=int64, numpy=\n",
|
||||
"array([[ 1, 45, 0, 0, 0, 0],\n",
|
||||
" [ 112, 1271, 1, 3, 1747, 158]], dtype=int64)>"
|
||||
]
|
||||
},
|
||||
"execution_count": 6,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"vectorizer(['Hello, world!','I am glad to meet you!'])"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"在這裡我們可以看到嵌入：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"array([[[ 1.53059261e-02, 6.80514947e-02, 3.14026810e-02, ...,\n",
|
||||
" -8.92002955e-02, 1.52911525e-04, -5.65562584e-02],\n",
|
||||
" [ 2.57456154e-01, 2.79364467e-01, -2.03605562e-01, ...,\n",
|
||||
" -2.07474351e-01, 8.31158683e-02, -2.03911960e-01],\n",
|
||||
" [ 3.98201384e-02, -8.03454965e-03, 2.39790026e-02, ...,\n",
|
||||
" -7.18549127e-04, 2.66963355e-02, -4.30646613e-02],\n",
|
||||
" [ 3.98201384e-02, -8.03454965e-03, 2.39790026e-02, ...,\n",
|
||||
" -7.18549127e-04, 2.66963355e-02, -4.30646613e-02],\n",
|
||||
" [ 3.98201384e-02, -8.03454965e-03, 2.39790026e-02, ...,\n",
|
||||
" -7.18549127e-04, 2.66963355e-02, -4.30646613e-02],\n",
|
||||
" [ 3.98201384e-02, -8.03454965e-03, 2.39790026e-02, ...,\n",
|
||||
" -7.18549127e-04, 2.66963355e-02, -4.30646613e-02]],\n",
|
||||
"\n",
|
||||
" [[ 1.89674050e-01, 2.61548996e-01, -3.67433839e-02, ...,\n",
|
||||
" -2.07366899e-01, -1.05442435e-01, -2.36952081e-01],\n",
|
||||
" [ 6.16133213e-02, 1.80511594e-01, 9.77298319e-02, ...,\n",
|
||||
" -5.46628237e-02, -1.07340455e-01, -1.06589928e-01],\n",
|
||||
" [ 1.53059261e-02, 6.80514947e-02, 3.14026810e-02, ...,\n",
|
||||
" -8.92002955e-02, 1.52911525e-04, -5.65562584e-02],\n",
|
||||
" [-4.84890305e-02, -8.41715634e-02, 1.51529670e-01, ...,\n",
|
||||
" 1.28192469e-01, -7.77286515e-02, 1.26041949e-01],\n",
|
||||
" [-4.17212099e-02, -5.60694858e-02, 4.08860669e-02, ...,\n",
|
||||
" 8.70475471e-02, 8.92383084e-02, 1.67974353e-01],\n",
|
||||
" [ 2.85779923e-01, 4.57767487e-01, 4.52292450e-02, ...,\n",
|
||||
" -1.97419018e-01, -2.04659685e-01, -2.79758364e-01]]],\n",
|
||||
" dtype=float32)"
|
||||
]
|
||||
},
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model.layers[1](vectorizer(['Hello, world!','I am glad to meet you!'])).numpy()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"> **注意**：為了減少填充的數量，在某些情況下，將數據集中所有序列按長度（更準確地說，是按標記數量）遞增的順序排序是有意義的。這將確保每個小批次包含長度相似的序列。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 語義嵌入：Word2Vec\n",
|
||||
"\n",
|
||||
"在之前的例子中，嵌入層學習如何將文字映射到向量表示，但這些表示並不具有語義上的意義。如果能學習一種向量表示，使得相似的詞或同義詞在某種向量距離（例如歐幾里得距離）上彼此接近，那就更好了。\n",
|
||||
"\n",
|
||||
"為了達到這個目的，我們需要使用像 [Word2Vec](https://en.wikipedia.org/wiki/Word2vec) 這樣的技術，在大量文本集合上預訓練嵌入模型。Word2Vec 基於兩種主要架構，用於生成詞的分佈式表示：\n",
|
||||
"\n",
|
||||
" - **連續詞袋模型** (CBoW)，我們訓練模型根據周圍的上下文來預測一個詞。給定 ngram $(W_{-2},W_{-1},W_0,W_1,W_2)$，模型的目標是根據 $(W_{-2},W_{-1},W_1,W_2)$ 預測 $W_0$。\n",
|
||||
" - **連續跳字模型** (Skip-Gram) 與 CBoW 相反。模型使用周圍的上下文詞窗口來預測當前的詞。\n",
|
||||
"\n",
|
||||
"CBoW 的速度較快，而 Skip-Gram 雖然較慢，但在表示不常見的詞方面效果更好。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"為了試驗基於 Google News 數據集預訓練的 Word2Vec 嵌入，我們可以使用 **gensim** 庫。以下是找到與「neural」最相似的詞的示例。\n",
|
||||
"\n",
|
||||
"> **注意：** 當你第一次創建詞向量時，下載它們可能需要一些時間！\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 8,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import gensim.downloader as api\n",
|
||||
"w2v = api.load('word2vec-google-news-300')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 12,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"neuronal -> 0.7804799675941467\n",
|
||||
"neurons -> 0.7326500415802002\n",
|
||||
"neural_circuits -> 0.7252851724624634\n",
|
||||
"neuron -> 0.7174385190010071\n",
|
||||
"cortical -> 0.6941086649894714\n",
|
||||
"brain_circuitry -> 0.6923246383666992\n",
|
||||
"synaptic -> 0.6699118614196777\n",
|
||||
"neural_circuitry -> 0.6638563275337219\n",
|
||||
"neurochemical -> 0.6555314064025879\n",
|
||||
"neuronal_activity -> 0.6531826257705688\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"for w,p in w2v.most_similar('neural'):\n",
|
||||
" print(f\"{w} -> {p}\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們還可以從該詞中提取向量嵌入，用於訓練分類模型。該嵌入有300個組件，但這裡為了清晰起見，我們僅顯示向量的前20個組件：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 13,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"array([ 0.01226807, 0.06225586, 0.10693359, 0.05810547, 0.23828125,\n",
|
||||
" 0.03686523, 0.05151367, -0.20703125, 0.01989746, 0.10058594,\n",
|
||||
" -0.03759766, -0.1015625 , -0.15820312, -0.08105469, -0.0390625 ,\n",
|
||||
" -0.05053711, 0.16015625, 0.2578125 , 0.10058594, -0.25976562],\n",
|
||||
" dtype=float32)"
|
||||
]
|
||||
},
|
||||
"execution_count": 13,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"w2v['play'][:20]"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"語義嵌入的偉大之處在於你可以根據語義操控向量編碼。例如，我們可以要求找到一個詞，其向量表示盡可能接近詞語*國王*和*女人*，並且盡可能遠離詞語*男人*：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 14,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"('queen', 0.7118192911148071)"
|
||||
]
|
||||
},
|
||||
"execution_count": 14,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"w2v.most_similar(positive=['king','woman'],negative=['man'])[0]"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"tags": []
|
||||
},
|
||||
"source": [
|
||||
"上面的例子使用了一些內部的 GenSym 魔法，但其基本邏輯其實相當簡單。關於嵌入的一個有趣之處是，你可以對嵌入向量執行正常的向量運算，而這將反映在詞語**意義**上的操作。上面的例子可以用向量運算來表達：我們計算出對應於 **KING-MAN+WOMAN** 的向量（對應詞語的向量表示執行 `+` 和 `-` 操作），然後在字典中找到最接近該向量的詞語：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 15,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"'queen'"
|
||||
]
|
||||
},
|
||||
"execution_count": 15,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"# get the vector corresponding to kind-man+woman\n",
|
||||
"qvec = w2v['king']-1.7*w2v['man']+1.7*w2v['woman']\n",
|
||||
"# find the index of the closest embedding vector \n",
|
||||
"d = np.sum((w2v.vectors-qvec)**2,axis=1)\n",
|
||||
"min_idx = np.argmin(d)\n",
|
||||
"# find the corresponding word\n",
|
||||
"w2v.index_to_key[min_idx]"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"> **注意**：我們需要在 *man* 和 *woman* 向量中加入一個小係數——試試移除它們，看看會發生什麼。\n",
|
||||
"\n",
|
||||
"為了找到最接近的向量，我們使用 TensorFlow 的工具來計算我們的向量與詞彙表中所有向量之間的距離向量，然後使用 `argmin` 找出最小距離的詞的索引。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"雖然 Word2Vec 是一種表達詞語語義的好方法，但它有許多缺點，包括以下幾點：\n",
|
||||
"\n",
|
||||
"* CBoW 和 skip-gram 模型都是**預測型嵌入**，它們只考慮局部上下文。Word2Vec 無法利用全局上下文。\n",
|
||||
"* Word2Vec 沒有考慮到詞語的**形態學**，即詞語的意義可能取決於詞的不同部分，例如詞根。\n",
|
||||
"\n",
|
||||
"**FastText** 嘗試克服第二個限制，並在 Word2Vec 的基礎上進一步改進，通過學習每個詞的向量表示以及詞內的字符 n-gram。這些表示的值在每次訓練步驟中會被平均成一個向量。雖然這為預訓練增加了大量額外的計算，但它使詞嵌入能夠編碼子詞資訊。\n",
|
||||
"\n",
|
||||
"另一種方法，**GloVe**，採用了與 Word2Vec 不同的詞嵌入方法，基於詞-上下文矩陣的分解。首先，它構建了一個大型矩陣，記錄詞語在不同上下文中的出現次數，然後嘗試以降低維度的方式表示該矩陣，同時最小化重建損失。\n",
|
||||
"\n",
|
||||
"gensim 庫支持這些詞嵌入方法，您可以通過修改上方的模型加載代碼來嘗試使用它們。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 在 Keras 中使用預訓練的嵌入\n",
|
||||
"\n",
|
||||
"我們可以修改上述範例，在嵌入層的矩陣中預先填入語義嵌入，例如 Word2Vec。預訓練嵌入的詞彙表和文本語料庫的詞彙表可能不一致，因此我們需要選擇其中一個。在這裡，我們探討兩種可能的選擇：使用分詞器的詞彙表，或者使用 Word2Vec 嵌入的詞彙表。\n",
|
||||
"\n",
|
||||
"### 使用分詞器的詞彙表\n",
|
||||
"\n",
|
||||
"當使用分詞器的詞彙表時，詞彙表中的某些詞會有對應的 Word2Vec 嵌入，而某些則會缺失。假設我們的詞彙表大小為 `vocab_size`，而 Word2Vec 嵌入向量的長度為 `embed_size`，那麼嵌入層將由一個形狀為 `vocab_size`$\\times$`embed_size` 的權重矩陣表示。我們將通過遍歷詞彙表來填充這個矩陣：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 9,
|
||||
"metadata": {
|
||||
"tags": []
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Embedding size: 300\n",
|
||||
"Populating matrix, this will take some time...Done, found 4551 words, 784 words missing\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"embed_size = len(w2v.get_vector('hello'))\n",
|
||||
"print(f'Embedding size: {embed_size}')\n",
|
||||
"\n",
|
||||
"vocab = vectorizer.get_vocabulary()\n",
|
||||
"W = np.zeros((vocab_size,embed_size))\n",
|
||||
"print('Populating matrix, this will take some time...',end='')\n",
|
||||
"found, not_found = 0,0\n",
|
||||
"for i,w in enumerate(vocab):\n",
|
||||
" try:\n",
|
||||
" W[i] = w2v.get_vector(w)\n",
|
||||
" found+=1\n",
|
||||
" except:\n",
|
||||
" # W[i] = np.random.normal(0.0,0.3,size=(embed_size,))\n",
|
||||
" not_found+=1\n",
|
||||
"\n",
|
||||
"print(f\"Done, found {found} words, {not_found} words missing\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"對於不在 Word2Vec 詞彙表中的詞語，我們可以選擇將它們設為零向量，或者生成一個隨機向量。\n",
|
||||
"\n",
|
||||
"現在我們可以定義一個帶有預訓練權重的嵌入層：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 10,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"emb = keras.layers.Embedding(vocab_size,embed_size,weights=[W],trainable=False)\n",
|
||||
"model = keras.models.Sequential([\n",
|
||||
" vectorizer, emb,\n",
|
||||
" keras.layers.Lambda(lambda x: tf.reduce_mean(x,axis=1)),\n",
|
||||
" keras.layers.Dense(4, activation='softmax')\n",
|
||||
"])"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": []
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 11,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"938/938 [==============================] - 10s 10ms/step - loss: 1.1075 - acc: 0.7822 - val_loss: 0.9134 - val_acc: 0.8175\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<keras.callbacks.History at 0x2220226ef10>"
|
||||
]
|
||||
},
|
||||
"execution_count": 11,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model.compile(loss='sparse_categorical_crossentropy',metrics=['acc'])\n",
|
||||
"model.fit(ds_train.map(tupelize).batch(batch_size),\n",
|
||||
" validation_data=ds_test.map(tupelize).batch(batch_size))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"> **注意**：請注意，我們在建立 `Embedding` 時設置了 `trainable=False`，這表示我們不會重新訓練 Embedding 層。這可能會導致準確度稍微降低，但能加快訓練速度。\n",
|
||||
"\n",
|
||||
"### 使用嵌入詞彙表\n",
|
||||
"\n",
|
||||
"之前方法的一個問題是，TextVectorization 和 Embedding 使用的詞彙表不同。為了解決這個問題，我們可以採用以下其中一種解決方案：\n",
|
||||
"* 重新訓練 Word2Vec 模型以適配我們的詞彙表。\n",
|
||||
"* 使用預訓練 Word2Vec 模型的詞彙表來載入我們的數據集。在載入數據集時，可以指定使用的詞彙表。\n",
|
||||
"\n",
|
||||
"第二種方法看起來更簡單，所以我們來實現它。首先，我們將使用 Word2Vec 嵌入中指定的詞彙表來建立一個 `TextVectorization` 層：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 12,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"vocab = list(w2v.vocab.keys())\n",
|
||||
"vectorizer = keras.layers.experimental.preprocessing.TextVectorization(input_shape=(1,))\n",
|
||||
"vectorizer.set_vocabulary(vocab)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"gensim 的詞嵌入庫包含一個方便的函數 `get_keras_embeddings`，它會自動為你創建相應的 Keras 嵌入層。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 13,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Epoch 1/5\n",
|
||||
"938/938 [==============================] - 20s 14ms/step - loss: 1.3377 - acc: 0.4978 - val_loss: 1.2995 - val_acc: 0.5647\n",
|
||||
"Epoch 2/5\n",
|
||||
"938/938 [==============================] - 10s 10ms/step - loss: 1.2587 - acc: 0.5722 - val_loss: 1.2339 - val_acc: 0.5842\n",
|
||||
"Epoch 3/5\n",
|
||||
"938/938 [==============================] - 10s 10ms/step - loss: 1.1980 - acc: 0.5884 - val_loss: 1.1826 - val_acc: 0.5954\n",
|
||||
"Epoch 4/5\n",
|
||||
"938/938 [==============================] - 12s 13ms/step - loss: 1.1503 - acc: 0.6002 - val_loss: 1.1417 - val_acc: 0.6018\n",
|
||||
"Epoch 5/5\n",
|
||||
"938/938 [==============================] - 11s 12ms/step - loss: 1.1120 - acc: 0.6097 - val_loss: 1.1083 - val_acc: 0.6104\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<keras.callbacks.History at 0x2220ccb81c0>"
|
||||
]
|
||||
},
|
||||
"execution_count": 13,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model = keras.models.Sequential([\n",
|
||||
" vectorizer, \n",
|
||||
" w2v.get_keras_embedding(train_embeddings=False),\n",
|
||||
" keras.layers.Lambda(lambda x: tf.reduce_mean(x,axis=1)),\n",
|
||||
" keras.layers.Dense(4, activation='softmax')\n",
|
||||
"])\n",
|
||||
"model.compile(loss='sparse_categorical_crossentropy',metrics=['acc'])\n",
|
||||
"model.fit(ds_train.map(tupelize).batch(128),validation_data=ds_test.map(tupelize).batch(128),epochs=5)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們未能看到更高準確度的其中一個原因是因為我們的數據集中的某些詞語在預訓練的GloVe詞彙表中缺失，因此它們基本上被忽略了。為了解決這個問題，我們可以基於我們的數據集訓練自己的嵌入。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 語境嵌入\n",
|
||||
"\n",
|
||||
"傳統預訓練嵌入表示（例如 Word2Vec）的一個主要限制是，雖然它們可以捕捉到某些詞語的含義，但無法區分不同的意思。這可能會在下游模型中引起問題。\n",
|
||||
"\n",
|
||||
"例如，單詞「play」在以下兩個句子中有不同的意思：\n",
|
||||
"- 我去劇院看了一場**戲劇**。\n",
|
||||
"- John 想和他的朋友一起**玩**。\n",
|
||||
"\n",
|
||||
"我們提到的預訓練嵌入會將「play」這個詞的兩種意思表示為相同的嵌入。為了解決這個限制，我們需要基於**語言模型**來構建嵌入，語言模型是在大量文本語料上訓練的，並且*知道*單詞如何在不同語境中組合使用。討論語境嵌入超出了本教程的範圍，但我們會在下一單元討論語言模型時回到這個主題。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件已使用人工智能翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。應以原文文件作為權威來源。對於關鍵資訊，建議尋求專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或誤釋不承擔責任。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"interpreter": {
|
||||
"hash": "0cb620c6d4b9f7a635928804c26cf22403d89d98d79684e4529119355ee6d5a5"
|
||||
},
|
||||
"kernel_info": {
|
||||
"name": "conda-env-py37_tensorflow-py"
|
||||
},
|
||||
"kernelspec": {
|
||||
"display_name": "py37_tensorflow",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.8.12"
|
||||
},
|
||||
"nteract": {
|
||||
"version": "nteract-front-end@1.0.0"
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "b859482be7f61d1eadc2c6a2720a37e4",
|
||||
"translation_date": "2025-08-31T10:53:44+00:00",
|
||||
"source_file": "lessons/5-NLP/14-Embeddings/EmbeddingsTF.ipynb",
|
||||
"language_code": "hk"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 4
|
||||
}
|
||||
|
|
@ -0,0 +1,576 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"id": "NXTSugt6ieXh"
|
||||
},
|
||||
"source": [
|
||||
"## 訓練 CBoW 模型\n",
|
||||
"\n",
|
||||
"此筆記本是 [AI for Beginners Curriculum](http://aka.ms/ai-beginners) 的一部分\n",
|
||||
"\n",
|
||||
"在這個例子中，我們將學習如何訓練 CBoW 語言模型，來獲取我們自己的 Word2Vec 嵌入空間。我們將使用 AG News 數據集作為文本來源。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"import torch\n",
|
||||
"import torchtext\n",
|
||||
"import os\n",
|
||||
"import collections\n",
|
||||
"import builtins\n",
|
||||
"import random\n",
|
||||
"import numpy as np"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "q-UiiJUKaxHj"
|
||||
},
|
||||
"execution_count": null,
|
||||
"outputs": []
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "TFbR8CZaTZ1q"
|
||||
},
|
||||
"execution_count": null,
|
||||
"outputs": []
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"source": [
|
||||
"首先讓我們載入數據集並定義分詞器和詞彙表。我們將 `vocab_size` 設置為 5000 以稍微限制計算量。\n"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "HIwC7lI5T-ov"
|
||||
}
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"def load_dataset(ngrams = 1, min_freq = 1, vocab_size = 5000 , lines_cnt = 500):\n",
|
||||
" tokenizer = torchtext.data.utils.get_tokenizer('basic_english')\n",
|
||||
" print(\"Loading dataset...\")\n",
|
||||
" test_dataset, train_dataset = torchtext.datasets.AG_NEWS(root='./data')\n",
|
||||
" train_dataset = list(train_dataset)\n",
|
||||
" test_dataset = list(test_dataset)\n",
|
||||
" classes = ['World', 'Sports', 'Business', 'Sci/Tech']\n",
|
||||
" print('Building vocab...')\n",
|
||||
" counter = collections.Counter()\n",
|
||||
" for i, (_, line) in enumerate(train_dataset):\n",
|
||||
" counter.update(torchtext.data.utils.ngrams_iterator(tokenizer(line),ngrams=ngrams))\n",
|
||||
" if i == lines_cnt:\n",
|
||||
" break\n",
|
||||
" vocab = torchtext.vocab.Vocab(collections.Counter(dict(counter.most_common(vocab_size))), min_freq=min_freq)\n",
|
||||
" return train_dataset, test_dataset, classes, vocab, tokenizer"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "wdZuygtgiuLG"
|
||||
},
|
||||
"execution_count": null,
|
||||
"outputs": []
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"train_dataset, test_dataset, _, vocab, tokenizer = load_dataset()"
|
||||
],
|
||||
"metadata": {
|
||||
"colab": {
|
||||
"base_uri": "https://localhost:8080/"
|
||||
},
|
||||
"id": "4d1nU1gsivGu",
|
||||
"outputId": "949fe272-ae0e-49f5-c373-6703458b3a74"
|
||||
},
|
||||
"execution_count": null,
|
||||
"outputs": [
|
||||
{
|
||||
"output_type": "stream",
|
||||
"name": "stdout",
|
||||
"text": [
|
||||
"Loading dataset...\n",
|
||||
"Building vocab...\n"
|
||||
]
|
||||
}
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"def encode(x, vocabulary, tokenizer = tokenizer):\n",
|
||||
" return [vocabulary[s] for s in tokenizer(x)]"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "1XDYNhG8ToFV"
|
||||
},
|
||||
"execution_count": null,
|
||||
"outputs": []
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"id": "LIlQk6_PaHVY"
|
||||
},
|
||||
"source": [
|
||||
"## CBoW 模型\n",
|
||||
"\n",
|
||||
"CBoW 學習根據 $2N$ 個相鄰的詞來預測一個詞。例如，當 $N=1$ 時，我們可以從句子 *I like to train networks* 中得到以下配對：(like, I)、(I, like)、(to, like)、(like, to)、(train, to)、(to, train)、(networks, train)、(train, networks)。在這裡，第一個詞是作為輸入的相鄰詞，第二個詞是我們要預測的詞。\n",
|
||||
"\n",
|
||||
"為了構建一個用來預測下一個詞的網絡，我們需要提供相鄰詞作為輸入，並獲得詞的編號作為輸出。CBoW 網絡的架構如下：\n",
|
||||
"\n",
|
||||
"* 輸入詞會通過嵌入層。這個嵌入層就是我們的 Word2Vec 嵌入，因此我們會將其單獨定義為 `embedder` 變量。在這個例子中，我們會使用嵌入大小為 30，儘管你可能想嘗試更高的維度（真實的 Word2Vec 通常是 300 維）。\n",
|
||||
"* 嵌入向量接著會傳遞到一個線性層，該層將預測輸出詞。因此它有 `vocab_size` 個神經元。\n",
|
||||
"\n",
|
||||
"對於輸出，如果我們使用 `CrossEntropyLoss` 作為損失函數，我們也需要僅提供詞的編號作為期望結果，而不需要使用 one-hot 編碼。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"vocab_size = len(vocab)\n",
|
||||
"\n",
|
||||
"embedder = torch.nn.Embedding(num_embeddings = vocab_size, embedding_dim = 30)\n",
|
||||
"model = torch.nn.Sequential(\n",
|
||||
" embedder,\n",
|
||||
" torch.nn.Linear(in_features = 30, out_features = vocab_size),\n",
|
||||
")\n",
|
||||
"\n",
|
||||
"print(model)"
|
||||
],
|
||||
"metadata": {
|
||||
"colab": {
|
||||
"base_uri": "https://localhost:8080/"
|
||||
},
|
||||
"id": "akKTcKQKkfl2",
|
||||
"outputId": "da687e3e-a8ec-4c1a-e456-ab8cd6ac7dad"
|
||||
},
|
||||
"execution_count": null,
|
||||
"outputs": [
|
||||
{
|
||||
"output_type": "stream",
|
||||
"name": "stdout",
|
||||
"text": [
|
||||
"Sequential(\n",
|
||||
" (0): Embedding(5002, 30)\n",
|
||||
" (1): Linear(in_features=30, out_features=5002, bias=True)\n",
|
||||
")\n"
|
||||
]
|
||||
}
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"id": "Nud6jgGPaHVa"
|
||||
},
|
||||
"source": [
|
||||
"## 準備訓練數據\n",
|
||||
"\n",
|
||||
"現在讓我們編寫主要函數，用於從文本中計算 CBoW 詞對。這個函數將允許我們指定窗口大小，並返回一組詞對——輸入詞和輸出詞。請注意，這個函數既可以用於詞，也可以用於向量/張量——這將使我們能夠在將文本傳遞給 `to_cbow` 函數之前對其進行編碼。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {
|
||||
"colab": {
|
||||
"base_uri": "https://localhost:8080/"
|
||||
},
|
||||
"id": "x-dsXygOieXn",
|
||||
"outputId": "c2218280-e540-40ba-9546-efe48d0d714f"
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"output_type": "stream",
|
||||
"name": "stdout",
|
||||
"text": [
|
||||
"[['like', 'I'], ['to', 'I'], ['I', 'like'], ['to', 'like'], ['train', 'like'], ['I', 'to'], ['like', 'to'], ['train', 'to'], ['networks', 'to'], ['like', 'train'], ['to', 'train'], ['networks', 'train'], ['to', 'networks'], ['train', 'networks']]\n",
|
||||
"[[232, 172], [5, 172], [172, 232], [5, 232], [0, 232], [172, 5], [232, 5], [0, 5], [1202, 5], [232, 0], [5, 0], [1202, 0], [5, 1202], [0, 1202]]\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"def to_cbow(sent,window_size=2):\n",
|
||||
" res = []\n",
|
||||
" for i,x in enumerate(sent):\n",
|
||||
" for j in range(max(0,i-window_size),min(i+window_size+1,len(sent))):\n",
|
||||
" if i!=j:\n",
|
||||
" res.append([sent[j],x])\n",
|
||||
" return res\n",
|
||||
"\n",
|
||||
"print(to_cbow(['I','like','to','train','networks']))\n",
|
||||
"print(to_cbow(encode('I like to train networks', vocab)))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"id": "XVaaDLjaaHVb"
|
||||
},
|
||||
"source": [
|
||||
"讓我們準備訓練數據集。我們將瀏覽所有新聞，調用 `to_cbow` 來獲取單詞對列表，並將這些對添加到 `X` 和 `Y` 中。為了節省時間，我們只考慮前 10k 條新聞項目——如果你有更多時間等待，並希望獲得更好的嵌入，可以輕鬆移除這個限制 :)\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {
|
||||
"id": "54b-Gd9TieXo"
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"X = []\n",
|
||||
"Y = []\n",
|
||||
"for i, x in zip(range(10000), train_dataset):\n",
|
||||
" for w1, w2 in to_cbow(encode(x[1], vocab), window_size = 5):\n",
|
||||
" X.append(w1)\n",
|
||||
" Y.append(w2)\n",
|
||||
"\n",
|
||||
"X = torch.tensor(X)\n",
|
||||
"Y = torch.tensor(Y)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"source": [
|
||||
"我們還會將該數據轉換為一個數據集，並創建數據加載器：\n"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "cwWy0PzXWhN5"
|
||||
}
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"class SimpleIterableDataset(torch.utils.data.IterableDataset):\n",
|
||||
" def __init__(self, X, Y):\n",
|
||||
" super(SimpleIterableDataset).__init__()\n",
|
||||
" self.data = []\n",
|
||||
" for i in range(len(X)):\n",
|
||||
" self.data.append( (Y[i], X[i]) )\n",
|
||||
" random.shuffle(self.data)\n",
|
||||
"\n",
|
||||
" def __iter__(self):\n",
|
||||
" return iter(self.data)"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "mfoAcGPFZU8p"
|
||||
},
|
||||
"execution_count": null,
|
||||
"outputs": []
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"id": "e4NQ_-5waHVc"
|
||||
},
|
||||
"source": [
|
||||
"我們還會將該數據轉換為一個數據集，並創建數據加載器：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {
|
||||
"id": "AbLUcojlieXo"
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"ds = SimpleIterableDataset(X, Y)\n",
|
||||
"dl = torch.utils.data.DataLoader(ds, batch_size = 256)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"id": "pKQr7sXeaHVc"
|
||||
},
|
||||
"source": [
|
||||
"現在讓我們進行實際訓練。我們將使用 `SGD` 優化器，並設定相當高的學習率。你也可以嘗試使用其他優化器，例如 `Adam`。我們將先訓練 10 個周期——如果你希望更低的損失，可以重新執行此單元格。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"def train_epoch(net, dataloader, lr = 0.01, optimizer = None, loss_fn = torch.nn.CrossEntropyLoss(), epochs = None, report_freq = 1):\n",
|
||||
" optimizer = optimizer or torch.optim.Adam(net.parameters(), lr = lr)\n",
|
||||
" loss_fn = loss_fn.to(device)\n",
|
||||
" net.train()\n",
|
||||
"\n",
|
||||
" for i in range(epochs):\n",
|
||||
" total_loss, j = 0, 0, \n",
|
||||
" for labels, features in dataloader:\n",
|
||||
" optimizer.zero_grad()\n",
|
||||
" features, labels = features.to(device), labels.to(device)\n",
|
||||
" out = net(features)\n",
|
||||
" loss = loss_fn(out, labels)\n",
|
||||
" loss.backward()\n",
|
||||
" optimizer.step()\n",
|
||||
" total_loss += loss\n",
|
||||
" j += 1\n",
|
||||
" if i % report_freq == 0:\n",
|
||||
" print(f\"Epoch: {i+1}: loss={total_loss.item()/j}\")\n",
|
||||
"\n",
|
||||
" return total_loss.item()/j"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "HeeCYKr_KF1w"
|
||||
},
|
||||
"execution_count": null,
|
||||
"outputs": []
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"train_epoch(net = model, dataloader = dl, optimizer = torch.optim.SGD(model.parameters(), lr = 0.1), loss_fn = torch.nn.CrossEntropyLoss(), epochs = 10)"
|
||||
],
|
||||
"metadata": {
|
||||
"colab": {
|
||||
"base_uri": "https://localhost:8080/"
|
||||
},
|
||||
"id": "KVgwGtDHgDlT",
|
||||
"outputId": "2447833f-f0e3-4566-c33d-addbfe2f451d"
|
||||
},
|
||||
"execution_count": null,
|
||||
"outputs": [
|
||||
{
|
||||
"output_type": "stream",
|
||||
"name": "stdout",
|
||||
"text": [
|
||||
"Epoch: 1: loss=5.664632366860172\n",
|
||||
"Epoch: 2: loss=5.632101973960962\n",
|
||||
"Epoch: 3: loss=5.610399051405015\n",
|
||||
"Epoch: 4: loss=5.594621561080262\n",
|
||||
"Epoch: 5: loss=5.582538017415446\n",
|
||||
"Epoch: 6: loss=5.572900234519603\n",
|
||||
"Epoch: 7: loss=5.564951676341915\n",
|
||||
"Epoch: 8: loss=5.558288112064614\n",
|
||||
"Epoch: 9: loss=5.552576955031129\n",
|
||||
"Epoch: 10: loss=5.547634165194347\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"output_type": "execute_result",
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"5.547634165194347"
|
||||
]
|
||||
},
|
||||
"metadata": {},
|
||||
"execution_count": 16
|
||||
}
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"id": "W8u2qXZmaHVd"
|
||||
},
|
||||
"source": [
|
||||
"## 嘗試使用 Word2Vec\n",
|
||||
"\n",
|
||||
"要使用 Word2Vec，我們先提取與詞彙表中所有單詞對應的向量：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {
|
||||
"id": "r8TatcXjkU_t"
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"vectors = torch.stack([embedder(torch.tensor(vocab[s])) for s in vocab.itos], 0)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"id": "3OcX21UOaHVd"
|
||||
},
|
||||
"source": [
|
||||
"讓我們看看，例如，單詞 **Paris** 是如何被編碼成一個向量的：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {
|
||||
"colab": {
|
||||
"base_uri": "https://localhost:8080/"
|
||||
},
|
||||
"id": "bz6tAeLzieXp",
|
||||
"outputId": "5b20850e-4342-45e9-f840-cfac2b4d61d8"
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"output_type": "stream",
|
||||
"name": "stdout",
|
||||
"text": [
|
||||
"tensor([-0.0915, 2.1224, -0.0281, -0.6819, 1.1219, 0.6458, -1.3704, -1.3314,\n",
|
||||
" -1.1437, 0.4496, 0.2301, -0.3515, -0.8485, 1.0481, 0.4386, -0.8949,\n",
|
||||
" 0.5644, 1.0939, -2.5096, 3.2949, -0.2601, -0.8640, 0.1421, -0.0804,\n",
|
||||
" -0.5083, -1.0560, 0.9753, -0.5949, -1.6046, 0.5774],\n",
|
||||
" grad_fn=<EmbeddingBackward>)\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"paris_vec = embedder(torch.tensor(vocab['paris']))\n",
|
||||
"print(paris_vec)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"id": "pHTJlaeYaHVd"
|
||||
},
|
||||
"source": [
|
||||
"使用 Word2Vec 來尋找同義詞是很有趣的。以下函數將返回與給定輸入最接近的 `n` 個詞。為了找到它們，我們計算 $|w_i - v|$ 的範數，其中 $v$ 是與我們輸入詞對應的向量，而 $w_i$ 是詞彙表中第 $i$ 個詞的編碼。我們然後對數組進行排序，並使用 `argsort` 返回相應的索引，接著取列表的前 `n` 個元素，這些元素編碼了詞彙表中最接近詞的位置。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {
|
||||
"colab": {
|
||||
"base_uri": "https://localhost:8080/"
|
||||
},
|
||||
"id": "NlZyi-_olFar",
|
||||
"outputId": "b5dbb163-88c4-4d5a-eaf2-6751f700e98c"
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"output_type": "execute_result",
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"['microsoft', 'quoted', 'lp', 'rate', 'top']"
|
||||
]
|
||||
},
|
||||
"metadata": {},
|
||||
"execution_count": 56
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"def close_words(x, n = 5):\n",
|
||||
" vec = embedder(torch.tensor(vocab[x]))\n",
|
||||
" top5 = np.linalg.norm(vectors.detach().numpy() - vec.detach().numpy(), axis = 1).argsort()[:n]\n",
|
||||
" return [ vocab.itos[x] for x in top5 ]\n",
|
||||
"\n",
|
||||
"close_words('microsoft')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {
|
||||
"colab": {
|
||||
"base_uri": "https://localhost:8080/"
|
||||
},
|
||||
"id": "-dQq7xeAln0U",
|
||||
"outputId": "66f768c3-c248-4bfd-ce4f-c8ffc6d0dd0d"
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"output_type": "execute_result",
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"['basketball', 'lot', 'sinai', 'states', 'healthdaynews']"
|
||||
]
|
||||
},
|
||||
"metadata": {},
|
||||
"execution_count": 51
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"close_words('basketball')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {
|
||||
"colab": {
|
||||
"base_uri": "https://localhost:8080/"
|
||||
},
|
||||
"id": "fJXqK26b29sa",
|
||||
"outputId": "78f0baba-ffd0-485a-dd87-0a12bedfd7fa"
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"output_type": "execute_result",
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"['funds', 'travel', 'sydney', 'japan', 'business']"
|
||||
]
|
||||
},
|
||||
"metadata": {},
|
||||
"execution_count": 77
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"close_words('funds')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"id": "My0VeTDd3Ji8"
|
||||
},
|
||||
"source": [
|
||||
"## 重點\n",
|
||||
"\n",
|
||||
"透過使用像 CBoW 這樣的巧妙技術，我們可以訓練 Word2Vec 模型。你也可以嘗試訓練 skip-gram 模型，該模型是基於給定中心詞來預測鄰近詞，看看它的表現如何。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件已使用人工智能翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。儘管我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。原始語言的文件應被視為權威來源。對於重要信息，建議使用專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或錯誤解釋不承擔責任。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"colab": {
|
||||
"collapsed_sections": [],
|
||||
"name": "CBoW-PyTorch.ipynb",
|
||||
"provenance": []
|
||||
},
|
||||
"interpreter": {
|
||||
"hash": "16af2a8bbb083ea23e5e41c7f5787656b2ce26968575d8763f2c4b17f9cd711f"
|
||||
},
|
||||
"kernelspec": {
|
||||
"display_name": "Python 3.8.12 ('py38')",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.8.12"
|
||||
},
|
||||
"orig_nbformat": 4,
|
||||
"gpuClass": "standard",
|
||||
"coopTranslator": {
|
||||
"original_hash": "36df28efe3fe40b6fb0a7fa48fe3ea82",
|
||||
"translation_date": "2025-08-31T10:36:54+00:00",
|
||||
"source_file": "lessons/5-NLP/15-LanguageModeling/CBoW-PyTorch.ipynb",
|
||||
"language_code": "hk"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 0
|
||||
}
|
||||
File diff suppressed because it is too large
Load Diff
|
|
@ -0,0 +1,479 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 循環神經網絡\n",
|
||||
"\n",
|
||||
"在上一個模組中，我們使用了豐富的語義表示來處理文本，並在嵌入層之上使用了一個簡單的線性分類器。這種架構能夠捕捉句子中詞語的聚合意義，但它並未考慮詞語的**順序**，因為嵌入層上的聚合操作已經將原始文本中的這種信息移除了。由於這些模型無法建模詞語的順序，因此它們無法解決更複雜或更具歧義的任務，例如文本生成或問答系統。\n",
|
||||
"\n",
|
||||
"為了捕捉文本序列的意義，我們需要使用另一種神經網絡架構，稱為**循環神經網絡**（Recurrent Neural Network，簡稱 RNN）。在 RNN 中，我們將句子逐個符號地傳遞給網絡，網絡會生成某種**狀態**，然後我們將該狀態與下一個符號一起再次傳遞給網絡。\n",
|
||||
"\n",
|
||||
"給定輸入的符號序列 $X_0,\\dots,X_n$，RNN 會創建一個神經網絡塊的序列，並通過反向傳播對這個序列進行端到端的訓練。每個網絡塊接收一對 $(X_i,S_i)$ 作為輸入，並生成 $S_{i+1}$ 作為結果。最終的狀態 $S_n$ 或輸出 $X_n$ 會進入線性分類器以生成結果。所有的網絡塊共享相同的權重，並通過一次反向傳播進行端到端的訓練。\n",
|
||||
"\n",
|
||||
"由於狀態向量 $S_0,\\dots,S_n$ 會在網絡中傳遞，因此它能夠學習詞語之間的順序依賴關係。例如，當單詞 *not* 出現在序列中的某處時，網絡可以學會在狀態向量中否定某些元素，從而實現否定的效果。\n",
|
||||
"\n",
|
||||
"> 由於圖片中所有 RNN 塊的權重是共享的，因此同一張圖片可以表示為一個帶有循環反饋迴路的單一塊（右側），該迴路將網絡的輸出狀態傳回到輸入。\n",
|
||||
"\n",
|
||||
"現在讓我們看看循環神經網絡如何幫助我們對新聞數據集進行分類。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 1,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Loading dataset...\n",
|
||||
"Building vocab...\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"import torch\n",
|
||||
"import torchtext\n",
|
||||
"from torchnlp import *\n",
|
||||
"train_dataset, test_dataset, classes, vocab = load_dataset()\n",
|
||||
"vocab_size = len(vocab)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 簡單 RNN 分類器\n",
|
||||
"\n",
|
||||
"在簡單 RNN 的情況下，每個循環單元是一個簡單的線性網絡，它接收連接的輸入向量和狀態向量，並生成一個新的狀態向量。PyTorch 使用 `RNNCell` 類來表示這個單元，而由這些單元組成的網絡則表示為 `RNN` 層。\n",
|
||||
"\n",
|
||||
"為了定義一個 RNN 分類器，我們首先會應用一個嵌入層來降低輸入詞彙的維度，然後在其上添加一個 RNN 層：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 2,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"class RNNClassifier(torch.nn.Module):\n",
|
||||
" def __init__(self, vocab_size, embed_dim, hidden_dim, num_class):\n",
|
||||
" super().__init__()\n",
|
||||
" self.hidden_dim = hidden_dim\n",
|
||||
" self.embedding = torch.nn.Embedding(vocab_size, embed_dim)\n",
|
||||
" self.rnn = torch.nn.RNN(embed_dim,hidden_dim,batch_first=True)\n",
|
||||
" self.fc = torch.nn.Linear(hidden_dim, num_class)\n",
|
||||
"\n",
|
||||
" def forward(self, x):\n",
|
||||
" batch_size = x.size(0)\n",
|
||||
" x = self.embedding(x)\n",
|
||||
" x,h = self.rnn(x)\n",
|
||||
" return self.fc(x.mean(dim=1))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"> **注意：** 為了簡化，我們在這裡使用未經訓練的嵌入層，但如果想要更好的效果，可以使用預訓練的嵌入層，例如 Word2Vec 或 GloVe 嵌入，這在之前的單元中已經描述過。為了更好地理解，您可能需要調整此程式碼以使用預訓練的嵌入。\n",
|
||||
"\n",
|
||||
"在我們的情況下，我們將使用填充的數據加載器，因此每個批次都會包含一些相同長度的填充序列。RNN 層將接收嵌入張量的序列，並產生兩個輸出：\n",
|
||||
"* $x$ 是每一步中 RNN 單元輸出的序列\n",
|
||||
"* $h$ 是序列最後一個元素的最終隱藏狀態\n",
|
||||
"\n",
|
||||
"接著，我們應用一個全連接的線性分類器來獲得類別數。\n",
|
||||
"\n",
|
||||
"> **注意：** RNN 的訓練相當困難，因為一旦 RNN 單元沿著序列長度展開，反向傳播所涉及的層數會非常多。因此，我們需要選擇較小的學習率，並在更大的數據集上訓練網絡以獲得良好的結果。這可能需要相當長的時間，因此建議使用 GPU。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {
|
||||
"scrolled": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"3200: acc=0.3090625\n",
|
||||
"6400: acc=0.38921875\n",
|
||||
"9600: acc=0.4590625\n",
|
||||
"12800: acc=0.511953125\n",
|
||||
"16000: acc=0.5506875\n",
|
||||
"19200: acc=0.57921875\n",
|
||||
"22400: acc=0.6070089285714285\n",
|
||||
"25600: acc=0.6304296875\n",
|
||||
"28800: acc=0.6484027777777778\n",
|
||||
"32000: acc=0.66509375\n",
|
||||
"35200: acc=0.6790056818181818\n",
|
||||
"38400: acc=0.6929166666666666\n",
|
||||
"41600: acc=0.7035817307692308\n",
|
||||
"44800: acc=0.7137276785714286\n",
|
||||
"48000: acc=0.72225\n",
|
||||
"51200: acc=0.73001953125\n",
|
||||
"54400: acc=0.7372794117647059\n",
|
||||
"57600: acc=0.7436631944444444\n",
|
||||
"60800: acc=0.7503947368421052\n",
|
||||
"64000: acc=0.75634375\n",
|
||||
"67200: acc=0.7615773809523809\n",
|
||||
"70400: acc=0.7662642045454545\n",
|
||||
"73600: acc=0.7708423913043478\n",
|
||||
"76800: acc=0.7751822916666666\n",
|
||||
"80000: acc=0.7790625\n",
|
||||
"83200: acc=0.7825\n",
|
||||
"86400: acc=0.7858564814814815\n",
|
||||
"89600: acc=0.7890513392857142\n",
|
||||
"92800: acc=0.7920474137931034\n",
|
||||
"96000: acc=0.7952708333333334\n",
|
||||
"99200: acc=0.7982258064516129\n",
|
||||
"102400: acc=0.80099609375\n",
|
||||
"105600: acc=0.8037594696969697\n",
|
||||
"108800: acc=0.8060569852941176\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=16, collate_fn=padify, shuffle=True)\n",
|
||||
"net = RNNClassifier(vocab_size,64,32,len(classes)).to(device)\n",
|
||||
"train_epoch(net,train_loader, lr=0.001)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 長短期記憶網絡 (LSTM)\n",
|
||||
"\n",
|
||||
"傳統 RNN 的主要問題之一是所謂的 **梯度消失** 問題。由於 RNN 是通過一次反向傳播從頭到尾進行訓練的，因此在將誤差傳遞到網絡的第一層時會遇到困難，導致網絡無法學習遠距離的詞元之間的關係。為了解決這個問題，可以通過使用所謂的 **門控機制** 引入 **顯式狀態管理**。其中最知名的兩種架構是 **長短期記憶網絡** (LSTM) 和 **門控循環單元** (GRU)。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"LSTM 網絡的結構與 RNN 類似，但有兩個狀態會從一層傳遞到下一層：實際狀態 $c$ 和隱藏向量 $h$。在每個單元中，隱藏向量 $h_i$ 與輸入 $x_i$ 拼接在一起，並通過 **門控機制** 控制狀態 $c$ 的變化。每個門控機制都是一個帶有 sigmoid 激活函數（輸出範圍為 $[0,1]$）的神經網絡，可以將其視為在與狀態向量相乘時的位掩碼。以下是這些門控機制（如上圖從左到右所示）：\n",
|
||||
"* **遺忘門** 接收隱藏向量並決定向量 $c$ 的哪些部分需要遺忘，哪些需要保留。\n",
|
||||
"* **輸入門** 從輸入和隱藏向量中提取一些信息，並將其插入到狀態中。\n",
|
||||
"* **輸出門** 通過帶有 $\\tanh$ 激活函數的線性層轉換狀態，然後使用隱藏向量 $h_i$ 選擇其部分組件以生成新的狀態 $c_{i+1}$。\n",
|
||||
"\n",
|
||||
"狀態 $c$ 的組件可以被視為一些可以開啟或關閉的標誌。例如，當我們在序列中遇到名字 *Alice* 時，我們可能會假設它指的是女性角色，並在狀態中設置一個標誌，表示句子中有一個女性名詞。當我們進一步遇到短語 *and Tom* 時，我們會設置一個標誌，表示句子中有複數名詞。因此，通過操控狀態，我們可以追蹤句子部分的語法屬性。\n",
|
||||
"\n",
|
||||
"> **Note**: 理解 LSTM 內部結構的一個很棒的資源是 Christopher Olah 的這篇文章 [Understanding LSTM Networks](https://colah.github.io/posts/2015-08-Understanding-LSTMs/)。\n",
|
||||
"\n",
|
||||
"雖然 LSTM 單元的內部結構看起來很複雜，但 PyTorch 將這些實現隱藏在 `LSTMCell` 類中，並提供了 `LSTM` 對象來表示整個 LSTM 層。因此，LSTM 分類器的實現將與我們之前看到的簡單 RNN 非常相似：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"class LSTMClassifier(torch.nn.Module):\n",
|
||||
" def __init__(self, vocab_size, embed_dim, hidden_dim, num_class):\n",
|
||||
" super().__init__()\n",
|
||||
" self.hidden_dim = hidden_dim\n",
|
||||
" self.embedding = torch.nn.Embedding(vocab_size, embed_dim)\n",
|
||||
" self.embedding.weight.data = torch.randn_like(self.embedding.weight.data)-0.5\n",
|
||||
" self.rnn = torch.nn.LSTM(embed_dim,hidden_dim,batch_first=True)\n",
|
||||
" self.fc = torch.nn.Linear(hidden_dim, num_class)\n",
|
||||
"\n",
|
||||
" def forward(self, x):\n",
|
||||
" batch_size = x.size(0)\n",
|
||||
" x = self.embedding(x)\n",
|
||||
" x,(h,c) = self.rnn(x)\n",
|
||||
" return self.fc(h[-1])"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": []
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"3200: acc=0.259375\n",
|
||||
"6400: acc=0.25859375\n",
|
||||
"9600: acc=0.26177083333333334\n",
|
||||
"12800: acc=0.2784375\n",
|
||||
"16000: acc=0.313\n",
|
||||
"19200: acc=0.3528645833333333\n",
|
||||
"22400: acc=0.3965625\n",
|
||||
"25600: acc=0.4385546875\n",
|
||||
"28800: acc=0.4752777777777778\n",
|
||||
"32000: acc=0.505375\n",
|
||||
"35200: acc=0.5326704545454546\n",
|
||||
"38400: acc=0.5557552083333334\n",
|
||||
"41600: acc=0.5760817307692307\n",
|
||||
"44800: acc=0.5954910714285714\n",
|
||||
"48000: acc=0.6118333333333333\n",
|
||||
"51200: acc=0.62681640625\n",
|
||||
"54400: acc=0.6404779411764706\n",
|
||||
"57600: acc=0.6520138888888889\n",
|
||||
"60800: acc=0.662828947368421\n",
|
||||
"64000: acc=0.673546875\n",
|
||||
"67200: acc=0.6831547619047619\n",
|
||||
"70400: acc=0.6917897727272727\n",
|
||||
"73600: acc=0.6997146739130434\n",
|
||||
"76800: acc=0.707109375\n",
|
||||
"80000: acc=0.714075\n",
|
||||
"83200: acc=0.7209134615384616\n",
|
||||
"86400: acc=0.727037037037037\n",
|
||||
"89600: acc=0.7326674107142858\n",
|
||||
"92800: acc=0.7379633620689655\n",
|
||||
"96000: acc=0.7433645833333333\n",
|
||||
"99200: acc=0.7479032258064516\n",
|
||||
"102400: acc=0.752119140625\n",
|
||||
"105600: acc=0.7562405303030303\n",
|
||||
"108800: acc=0.76015625\n",
|
||||
"112000: acc=0.7641339285714286\n",
|
||||
"115200: acc=0.7677777777777778\n",
|
||||
"118400: acc=0.7711233108108108\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"(0.03487814127604167, 0.7728)"
|
||||
]
|
||||
},
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"net = LSTMClassifier(vocab_size,64,32,len(classes)).to(device)\n",
|
||||
"train_epoch(net,train_loader, lr=0.001)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 打包序列\n",
|
||||
"\n",
|
||||
"在我們的例子中，我們需要用零向量填充小批量中的所有序列。雖然這會導致一些記憶體浪費，但對於 RNN 而言，更關鍵的是為填充的輸入項目創建了額外的 RNN 單元，這些單元參與了訓練，但並未攜帶任何重要的輸入資訊。如果能僅針對實際的序列長度來訓練 RNN，效果會更好。\n",
|
||||
"\n",
|
||||
"為此，PyTorch 引入了一種特殊格式來存儲填充的序列。假設我們有一個填充過的小批量輸入，看起來像這樣：\n",
|
||||
"```\n",
|
||||
"[[1,2,3,4,5],\n",
|
||||
" [6,7,8,0,0],\n",
|
||||
" [9,0,0,0,0]]\n",
|
||||
"```\n",
|
||||
"這裡的 0 代表填充的值，而輸入序列的實際長度向量是 `[5,3,1]`。\n",
|
||||
"\n",
|
||||
"為了有效地用填充序列訓練 RNN，我們希望先用較大的小批量（`[1,6,9]`）開始訓練第一組 RNN 單元，然後結束第三個序列的處理，並繼續用較小的小批量（`[2,7]`，`[3,8]`）進行訓練，依此類推。因此，打包序列被表示為一個向量——在我們的例子中是 `[1,6,9,2,7,3,8,4,5]`，以及長度向量（`[5,3,1]`），我們可以根據這些輕鬆重建原始的填充小批量。\n",
|
||||
"\n",
|
||||
"要生成打包序列，我們可以使用 `torch.nn.utils.rnn.pack_padded_sequence` 函數。所有的循環層，包括 RNN、LSTM 和 GRU，都支持將打包序列作為輸入，並生成打包輸出，這些輸出可以通過 `torch.nn.utils.rnn.pad_packed_sequence` 解碼。\n",
|
||||
"\n",
|
||||
"為了能夠生成打包序列，我們需要將長度向量傳遞給網絡，因此我們需要一個不同的函數來準備小批量：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 6,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"def pad_length(b):\n",
|
||||
" # build vectorized sequence\n",
|
||||
" v = [encode(x[1]) for x in b]\n",
|
||||
" # compute max length of a sequence in this minibatch and length sequence itself\n",
|
||||
" len_seq = list(map(len,v))\n",
|
||||
" l = max(len_seq)\n",
|
||||
" return ( # tuple of three tensors - labels, padded features, length sequence\n",
|
||||
" torch.LongTensor([t[0]-1 for t in b]),\n",
|
||||
" torch.stack([torch.nn.functional.pad(torch.tensor(t),(0,l-len(t)),mode='constant',value=0) for t in v]),\n",
|
||||
" torch.tensor(len_seq)\n",
|
||||
" )\n",
|
||||
"\n",
|
||||
"train_loader_len = torch.utils.data.DataLoader(train_dataset, batch_size=16, collate_fn=pad_length, shuffle=True)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"實際的網絡會與上面的 `LSTMClassifier` 非常相似，但在 `forward` 傳遞時，會同時接收填充過的小批量數據以及序列長度的向量。在計算嵌入後，我們會計算打包序列，將其傳遞給 LSTM 層，然後再將結果解包回來。\n",
|
||||
"\n",
|
||||
"> **注意**：我們實際上並沒有使用解包後的結果 `x`，因為在後續的計算中，我們使用的是隱藏層的輸出。因此，我們可以完全移除這段代碼中的解包部分。我們之所以將其放在這裡，是為了方便你在需要使用網絡輸出進行進一步計算時，能夠輕鬆修改這段代碼。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"class LSTMPackClassifier(torch.nn.Module):\n",
|
||||
" def __init__(self, vocab_size, embed_dim, hidden_dim, num_class):\n",
|
||||
" super().__init__()\n",
|
||||
" self.hidden_dim = hidden_dim\n",
|
||||
" self.embedding = torch.nn.Embedding(vocab_size, embed_dim)\n",
|
||||
" self.embedding.weight.data = torch.randn_like(self.embedding.weight.data)-0.5\n",
|
||||
" self.rnn = torch.nn.LSTM(embed_dim,hidden_dim,batch_first=True)\n",
|
||||
" self.fc = torch.nn.Linear(hidden_dim, num_class)\n",
|
||||
"\n",
|
||||
" def forward(self, x, lengths):\n",
|
||||
" batch_size = x.size(0)\n",
|
||||
" x = self.embedding(x)\n",
|
||||
" pad_x = torch.nn.utils.rnn.pack_padded_sequence(x,lengths,batch_first=True,enforce_sorted=False)\n",
|
||||
" pad_x,(h,c) = self.rnn(pad_x)\n",
|
||||
" x, _ = torch.nn.utils.rnn.pad_packed_sequence(pad_x,batch_first=True)\n",
|
||||
" return self.fc(h[-1])"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": []
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 8,
|
||||
"metadata": {
|
||||
"scrolled": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"3200: acc=0.285625\n",
|
||||
"6400: acc=0.33359375\n",
|
||||
"9600: acc=0.3876041666666667\n",
|
||||
"12800: acc=0.44078125\n",
|
||||
"16000: acc=0.4825\n",
|
||||
"19200: acc=0.5235416666666667\n",
|
||||
"22400: acc=0.5559821428571429\n",
|
||||
"25600: acc=0.58609375\n",
|
||||
"28800: acc=0.6116666666666667\n",
|
||||
"32000: acc=0.63340625\n",
|
||||
"35200: acc=0.6525284090909091\n",
|
||||
"38400: acc=0.668515625\n",
|
||||
"41600: acc=0.6822596153846154\n",
|
||||
"44800: acc=0.6948214285714286\n",
|
||||
"48000: acc=0.7052708333333333\n",
|
||||
"51200: acc=0.71521484375\n",
|
||||
"54400: acc=0.7239889705882353\n",
|
||||
"57600: acc=0.7315277777777778\n",
|
||||
"60800: acc=0.7388486842105263\n",
|
||||
"64000: acc=0.74571875\n",
|
||||
"67200: acc=0.7518303571428572\n",
|
||||
"70400: acc=0.7576988636363636\n",
|
||||
"73600: acc=0.7628940217391305\n",
|
||||
"76800: acc=0.7681510416666667\n",
|
||||
"80000: acc=0.7728125\n",
|
||||
"83200: acc=0.7772235576923077\n",
|
||||
"86400: acc=0.7815393518518519\n",
|
||||
"89600: acc=0.7857700892857142\n",
|
||||
"92800: acc=0.7895043103448276\n",
|
||||
"96000: acc=0.7930520833333333\n",
|
||||
"99200: acc=0.7959072580645161\n",
|
||||
"102400: acc=0.798994140625\n",
|
||||
"105600: acc=0.802064393939394\n",
|
||||
"108800: acc=0.8051378676470589\n",
|
||||
"112000: acc=0.8077857142857143\n",
|
||||
"115200: acc=0.8104600694444445\n",
|
||||
"118400: acc=0.8128293918918919\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"(0.029785829671223958, 0.8138166666666666)"
|
||||
]
|
||||
},
|
||||
"execution_count": 8,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"net = LSTMPackClassifier(vocab_size,64,32,len(classes)).to(device)\n",
|
||||
"train_epoch_emb(net,train_loader_len, lr=0.001,use_pack_sequence=True)\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"> **注意：** 您可能已經注意到我們傳遞給訓練函數的參數 `use_pack_sequence`。目前，`pack_padded_sequence` 函數要求長度序列張量位於 CPU 設備上，因此訓練函數需要避免在訓練時將長度序列數據移動到 GPU。您可以查看 [`torchnlp.py`](../../../../../lessons/5-NLP/16-RNN/torchnlp.py) 文件中 `train_emb` 函數的實現。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 雙向和多層 RNN\n",
|
||||
"\n",
|
||||
"在我們的例子中，所有的循環神經網絡都是單向運行的，從序列的開頭到結尾。這看起來很自然，因為它類似於我們閱讀和聆聽語音的方式。然而，在許多實際情況下，我們可以隨機訪問輸入序列，因此在兩個方向上運行循環計算可能更有意義。這樣的網絡被稱為 **雙向** RNN，可以通過在 RNN/LSTM/GRU 構造函數中傳遞參數 `bidirectional=True` 來創建。\n",
|
||||
"\n",
|
||||
"處理雙向網絡時，我們需要兩個隱藏狀態向量，每個方向一個。PyTorch 將這些向量編碼為一個大小是原來兩倍的向量，這非常方便，因為通常我們會將生成的隱藏狀態傳遞給全連接線性層，只需在創建該層時考慮到這個大小的增加即可。\n",
|
||||
"\n",
|
||||
"無論是單向還是雙向的循環網絡，都能在序列中捕捉某些模式，並將其存儲到狀態向量中或傳遞到輸出中。與卷積網絡類似，我們可以在第一層之上構建另一個循環層，以捕捉更高層次的模式，這些模式是由第一層提取的低層次模式構建而成的。這引出了 **多層 RNN** 的概念，它由兩層或更多的循環網絡組成，其中前一層的輸出作為下一層的輸入。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"*圖片來源於 [這篇精彩的文章](https://towardsdatascience.com/from-a-lstm-cell-to-a-multilayer-lstm-network-with-pytorch-2899eb5696f3) 作者 Fernando López*\n",
|
||||
"\n",
|
||||
"PyTorch 讓構建這樣的網絡變得非常簡單，因為只需在 RNN/LSTM/GRU 構造函數中傳遞參數 `num_layers`，即可自動構建多層循環網絡。這也意味著隱藏/狀態向量的大小會按比例增加，處理循環層的輸出時需要考慮到這一點。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## RNNs 用於其他任務\n",
|
||||
"\n",
|
||||
"在本單元中，我們已經看到 RNNs 可以用於序列分類，但事實上，它們還能處理更多任務，例如文本生成、機器翻譯等等。我們會在下一單元探討這些任務。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件已使用人工智能翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。儘管我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。原始語言的文件應被視為權威來源。對於重要資訊，建議使用專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或錯誤解釋概不負責。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"interpreter": {
|
||||
"hash": "16af2a8bbb083ea23e5e41c7f5787656b2ce26968575d8763f2c4b17f9cd711f"
|
||||
},
|
||||
"kernelspec": {
|
||||
"display_name": "Python 3.8.12 ('py38')",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.8.12"
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "522ee52ae3d5ae933e283286254e9a55",
|
||||
"translation_date": "2025-08-31T10:50:36+00:00",
|
||||
"source_file": "lessons/5-NLP/16-RNN/RNNPyTorch.ipynb",
|
||||
"language_code": "hk"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 2
|
||||
}
|
||||
|
|
@ -0,0 +1,460 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 循環神經網絡\n",
|
||||
"\n",
|
||||
"在上一個模組中，我們探討了文本的豐富語義表示。我們使用的架構能夠捕捉句子中詞語的聚合意義，但它並未考慮詞語的**順序**，因為嵌入後的聚合操作會移除原始文本中的這些信息。由於這些模型無法表示詞語的排列順序，因此它們無法解決更複雜或更具歧義的任務，例如文本生成或問題回答。\n",
|
||||
"\n",
|
||||
"為了捕捉文本序列的意義，我們將使用一種名為**循環神經網絡**（Recurrent Neural Network，簡稱 RNN）的神經網絡架構。在使用 RNN 時，我們會將句子逐個標記傳遞給網絡，網絡會生成某種**狀態**，然後我們將該狀態與下一個標記一起再次傳遞給網絡。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"給定標記的輸入序列 $X_0,\\dots,X_n$，RNN 會創建一系列神經網絡塊，並通過反向傳播對這個序列進行端到端的訓練。每個網絡塊以 $(X_i,S_i)$ 作為輸入，並生成 $S_{i+1}$ 作為結果。最終狀態 $S_n$ 或輸出 $Y_n$ 會進入線性分類器以生成結果。所有網絡塊共享相同的權重，並通過一次反向傳播進行端到端訓練。\n",
|
||||
"\n",
|
||||
"> 上圖展示了循環神經網絡的展開形式（左側）和更緊湊的循環表示形式（右側）。需要注意的是，所有 RNN 單元都具有相同的**可共享權重**。\n",
|
||||
"\n",
|
||||
"由於狀態向量 $S_0,\\dots,S_n$ 是通過網絡傳遞的，RNN 能夠學習詞語之間的順序依賴。例如，當某個序列中出現詞語 *not* 時，它可以學習在狀態向量中否定某些元素。\n",
|
||||
"\n",
|
||||
"在內部，每個 RNN 單元包含兩個權重矩陣：$W_H$ 和 $W_I$，以及偏置 $b$。在每個 RNN 步驟中，給定輸入 $X_i$ 和輸入狀態 $S_i$，輸出狀態的計算公式為 $S_{i+1} = f(W_H\\times S_i + W_I\\times X_i+b)$，其中 $f$ 是激活函數（通常為 $\\tanh$）。\n",
|
||||
"\n",
|
||||
"> 對於像文本生成（我們會在下一單元中探討）或機器翻譯這類問題，我們還希望在每個 RNN 步驟中獲得某些輸出值。在這種情況下，還會有另一個矩陣 $W_O$，輸出值的計算公式為 $Y_i=f(W_O\\times S_i+b_O)$。\n",
|
||||
"\n",
|
||||
"現在讓我們看看循環神經網絡如何幫助我們對新聞數據集進行分類。\n",
|
||||
"\n",
|
||||
"> 在沙盒環境中，我們需要運行以下單元格以確保所需的庫已安裝並且數據已預取。如果您在本地運行，可以跳過以下單元格。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 1,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import sys\n",
|
||||
"!{sys.executable} -m pip install --quiet tensorflow_datasets==4.4.0\n",
|
||||
"!cd ~ && wget -q -O - https://mslearntensorflowlp.blob.core.windows.net/data/tfds-ag-news.tgz | tar xz"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 2,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import tensorflow as tf\n",
|
||||
"from tensorflow import keras\n",
|
||||
"import tensorflow_datasets as tfds\n",
|
||||
"import numpy as np\n",
|
||||
"\n",
|
||||
"# We are going to be training pretty large models. In order not to face errors, we need\n",
|
||||
"# to set tensorflow option to grow GPU memory allocation when required\n",
|
||||
"physical_devices = tf.config.list_physical_devices('GPU') \n",
|
||||
"if len(physical_devices)>0:\n",
|
||||
" tf.config.experimental.set_memory_growth(physical_devices[0], True)\n",
|
||||
"\n",
|
||||
"ds_train, ds_test = tfds.load('ag_news_subset').values()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"nteract": {
|
||||
"transient": {
|
||||
"deleting": false
|
||||
}
|
||||
}
|
||||
},
|
||||
"source": [
|
||||
"在訓練大型模型時，GPU 記憶體分配可能會成為一個問題。我們可能需要嘗試不同的迷你批次大小，以確保數據能夠適配 GPU 記憶體，同時訓練速度也足夠快。如果你在自己的 GPU 機器上運行此代碼，可以嘗試調整迷你批次大小來加快訓練速度。\n",
|
||||
"\n",
|
||||
"> **Note**: 某些版本的 NVidia 驅動程式已知在訓練模型後不會釋放記憶體。我們在這個筆記本中運行了幾個範例，這可能會導致某些配置下記憶體耗盡，特別是當你在同一個筆記本中進行自己的實驗時。如果在開始訓練模型時遇到一些奇怪的錯誤，你可能需要重新啟動筆記本的內核。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {
|
||||
"collapsed": true,
|
||||
"jupyter": {
|
||||
"outputs_hidden": false,
|
||||
"source_hidden": false
|
||||
},
|
||||
"nteract": {
|
||||
"transient": {
|
||||
"deleting": false
|
||||
}
|
||||
}
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"batch_size = 16\n",
|
||||
"embed_size = 64"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 簡單 RNN 分類器\n",
|
||||
"\n",
|
||||
"在簡單 RNN 的情況下，每個循環單元都是一個簡單的線性網絡，它接收輸入向量和狀態向量，並生成新的狀態向量。在 Keras 中，這可以用 `SimpleRNN` 層來表示。\n",
|
||||
"\n",
|
||||
"雖然我們可以直接將 one-hot 編碼的標記傳遞給 RNN 層，但這並不是一個好主意，因為它們的維度太高。因此，我們會使用嵌入層來降低詞向量的維度，接著是 RNN 層，最後是一個 `Dense` 分類器。\n",
|
||||
"\n",
|
||||
"> **注意**：在維度不是很高的情況下，例如使用字符級標記化時，直接將 one-hot 編碼的標記傳遞給 RNN 單元可能是合理的。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Model: \"sequential\"\n",
|
||||
"_________________________________________________________________\n",
|
||||
"Layer (type) Output Shape Param # \n",
|
||||
"=================================================================\n",
|
||||
"text_vectorization (TextVect (None, None) 0 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"embedding (Embedding) (None, None, 64) 1280000 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"simple_rnn (SimpleRNN) (None, 16) 1296 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"dense (Dense) (None, 4) 68 \n",
|
||||
"=================================================================\n",
|
||||
"Total params: 1,281,364\n",
|
||||
"Trainable params: 1,281,364\n",
|
||||
"Non-trainable params: 0\n",
|
||||
"_________________________________________________________________\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"vocab_size = 20000\n",
|
||||
"\n",
|
||||
"vectorizer = keras.layers.experimental.preprocessing.TextVectorization(\n",
|
||||
" max_tokens=vocab_size,\n",
|
||||
" input_shape=(1,))\n",
|
||||
"\n",
|
||||
"model = keras.models.Sequential([\n",
|
||||
" vectorizer,\n",
|
||||
" keras.layers.Embedding(vocab_size, embed_size),\n",
|
||||
" keras.layers.SimpleRNN(16),\n",
|
||||
" keras.layers.Dense(4,activation='softmax')\n",
|
||||
"])\n",
|
||||
"\n",
|
||||
"model.summary()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"> **注意：** 我們在這裡使用未經訓練的嵌入層以簡化操作，但為了獲得更好的效果，我們可以使用 Word2Vec 預訓練的嵌入層，如前一單元所述。這會是一個很好的練習，讓你嘗試將此程式碼改為使用預訓練的嵌入層。\n",
|
||||
"\n",
|
||||
"現在讓我們開始訓練 RNN。一般來說，RNN 的訓練相當困難，因為當 RNN 單元沿著序列長度展開時，反向傳播所涉及的層數會非常多。因此，我們需要選擇較小的學習率，並在更大的數據集上訓練網絡以獲得良好的結果。這可能需要相當長的時間，因此建議使用 GPU。\n",
|
||||
"\n",
|
||||
"為了加快速度，我們將僅使用新聞標題來訓練 RNN 模型，省略描述部分。你可以嘗試加入描述進行訓練，看看是否能讓模型成功訓練。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {
|
||||
"scrolled": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Training vectorizer\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"def extract_title(x):\n",
|
||||
" return x['title']\n",
|
||||
"\n",
|
||||
"def tupelize_title(x):\n",
|
||||
" return (extract_title(x),x['label'])\n",
|
||||
"\n",
|
||||
"print('Training vectorizer')\n",
|
||||
"vectorizer.adapt(ds_train.take(2000).map(extract_title))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 6,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"7500/7500 [==============================] - 82s 11ms/step - loss: 0.6629 - acc: 0.7623 - val_loss: 0.5559 - val_acc: 0.7995\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<tensorflow.python.keras.callbacks.History at 0x7f3e0030d350>"
|
||||
]
|
||||
},
|
||||
"execution_count": 6,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model.compile(loss='sparse_categorical_crossentropy',metrics=['acc'], optimizer='adam')\n",
|
||||
"model.fit(ds_train.map(tupelize_title).batch(batch_size),validation_data=ds_test.map(tupelize_title).batch(batch_size))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"nteract": {
|
||||
"transient": {
|
||||
"deleting": false
|
||||
}
|
||||
}
|
||||
},
|
||||
"source": [
|
||||
"> **注意**，由於我們僅針對新聞標題進行訓練，因此準確性可能會較低。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 重溫變數序列\n",
|
||||
"\n",
|
||||
"記住，`TextVectorization` 層會自動在小批量中用填充標記來填充可變長度的序列。然而，這些填充標記也會參與訓練，並可能使模型的收斂變得更複雜。\n",
|
||||
"\n",
|
||||
"我們可以採取幾種方法來減少填充的數量。其中一種方法是根據序列長度重新排列數據集，並按大小分組所有序列。這可以使用 `tf.data.experimental.bucket_by_sequence_length` 函數來完成（請參閱[文件](https://www.tensorflow.org/api_docs/python/tf/data/experimental/bucket_by_sequence_length)）。\n",
|
||||
"\n",
|
||||
"另一種方法是使用**遮罩**。在 Keras 中，某些層支持額外的輸入，該輸入顯示哪些標記應在訓練時被考慮。要將遮罩整合到我們的模型中，我們可以選擇添加一個單獨的 `Masking` 層（[文件](https://keras.io/api/layers/core_layers/masking/)），或者在我們的 `Embedding` 層中指定 `mask_zero=True` 參數。\n",
|
||||
"\n",
|
||||
"> **Note**: 完成整個數據集的一個訓練週期大約需要 5 分鐘。如果你失去耐心，可以隨時中斷訓練。你也可以通過在 `ds_train` 和 `ds_test` 數據集後添加 `.take(...)` 子句來限制用於訓練的數據量。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"7500/7500 [==============================] - 371s 49ms/step - loss: 0.5401 - acc: 0.8079 - val_loss: 0.3780 - val_acc: 0.8822\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<tensorflow.python.keras.callbacks.History at 0x7f3dec118850>"
|
||||
]
|
||||
},
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"def extract_text(x):\n",
|
||||
" return x['title']+' '+x['description']\n",
|
||||
"\n",
|
||||
"def tupelize(x):\n",
|
||||
" return (extract_text(x),x['label'])\n",
|
||||
"\n",
|
||||
"model = keras.models.Sequential([\n",
|
||||
" vectorizer,\n",
|
||||
" keras.layers.Embedding(vocab_size,embed_size,mask_zero=True),\n",
|
||||
" keras.layers.SimpleRNN(16),\n",
|
||||
" keras.layers.Dense(4,activation='softmax')\n",
|
||||
"])\n",
|
||||
"\n",
|
||||
"model.compile(loss='sparse_categorical_crossentropy',metrics=['acc'], optimizer='adam')\n",
|
||||
"model.fit(ds_train.map(tupelize).batch(batch_size),validation_data=ds_test.map(tupelize).batch(batch_size))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在我們使用遮罩技術，可以在整個標題和描述的數據集上訓練模型。\n",
|
||||
"\n",
|
||||
"> **注意**：你有沒有注意到，我們一直在使用基於新聞標題訓練的向量化工具，而不是整篇文章的內容？這可能會導致某些詞元被忽略，因此重新訓練向量化工具會更好。不過，這可能只會帶來非常小的影響，所以為了簡化流程，我們會繼續使用之前預訓練的向量化工具。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## LSTM：長短期記憶\n",
|
||||
"\n",
|
||||
"RNNs 的主要問題之一是 **梯度消失**。RNNs 可以非常長，在反向傳播過程中可能難以將梯度傳遞回網絡的第一層。當出現這種情況時，網絡無法學習遠距離的標記之間的關係。解決這個問題的一種方法是通過使用 **閘門** 引入 **顯式狀態管理**。最常見的引入閘門的架構是 **長短期記憶**（LSTM）和 **門控循環單元**（GRU）。我們在這裡將討論 LSTM。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"LSTM 網絡的組織方式與 RNN 類似，但有兩個狀態會從層到層傳遞：實際狀態 $c$ 和隱藏向量 $h$。在每個單元中，隱藏向量 $h_{t-1}$ 與輸入 $x_t$ 結合，並共同控制狀態 $c_t$ 和輸出 $h_{t}$ 的變化，這是通過 **閘門** 完成的。每個閘門都有 sigmoid 激活函數（輸出範圍為 $[0,1]$），可以被視為在與狀態向量相乘時的位掩碼。LSTM 包含以下閘門（如上圖所示，從左到右）：\n",
|
||||
"* **遺忘閘門**：決定向量 $c_{t-1}$ 的哪些部分需要遺忘，哪些需要保留。\n",
|
||||
"* **輸入閘門**：決定來自輸入向量和前一隱藏向量的信息有多少應該被整合到狀態向量中。\n",
|
||||
"* **輸出閘門**：接收新的狀態向量並決定其哪些部分將用於生成新的隱藏向量 $h_t$。\n",
|
||||
"\n",
|
||||
"狀態 $c$ 的組件可以被視為可以開啟或關閉的標誌。例如，當我們在序列中遇到名字 *Alice* 時，我們猜測它指的是一位女性，並在狀態中啟動表示句子中有女性名詞的標誌。當我們進一步遇到 *and Tom* 這些詞時，我們會啟動表示句子中有複數名詞的標誌。因此，通過操控狀態，我們可以追蹤句子的語法屬性。\n",
|
||||
"\n",
|
||||
"> **注意**：這裡有一個很棒的資源可以幫助理解 LSTM 的內部結構：[理解 LSTM 網絡](https://colah.github.io/posts/2015-08-Understanding-LSTMs/) 作者是 Christopher Olah。\n",
|
||||
"\n",
|
||||
"雖然 LSTM 單元的內部結構看起來可能很複雜，但 Keras 將這些實現隱藏在 `LSTM` 層中，因此在上面的示例中，我們唯一需要做的就是替換循環層：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 8,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"15000/15000 [==============================] - 188s 13ms/step - loss: 0.5692 - acc: 0.7916 - val_loss: 0.3441 - val_acc: 0.8870\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<tensorflow.python.keras.callbacks.History at 0x7f3d6af5c350>"
|
||||
]
|
||||
},
|
||||
"execution_count": 8,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model = keras.models.Sequential([\n",
|
||||
" vectorizer,\n",
|
||||
" keras.layers.Embedding(vocab_size, embed_size),\n",
|
||||
" keras.layers.LSTM(8),\n",
|
||||
" keras.layers.Dense(4,activation='softmax')\n",
|
||||
"])\n",
|
||||
"\n",
|
||||
"model.compile(loss='sparse_categorical_crossentropy',metrics=['acc'], optimizer='adam')\n",
|
||||
"model.fit(ds_train.map(tupelize).batch(8),validation_data=ds_test.map(tupelize).batch(8))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": []
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 雙向和多層 RNN\n",
|
||||
"\n",
|
||||
"在我們之前的例子中，循環神經網絡（RNN）都是從序列的開頭運行到結尾。這對我們來說很自然，因為它符合我們閱讀或聆聽語音的方向。然而，對於需要隨機訪問輸入序列的情境來說，讓循環計算在兩個方向上運行會更有意義。允許雙向計算的 RNN 被稱為 **雙向 RNN**，可以通過將循環層包裹在一個特殊的 `Bidirectional` 層中來創建。\n",
|
||||
"\n",
|
||||
"> **Note**: `Bidirectional` 層會在內部生成該層的兩個副本，並將其中一個副本的 `go_backwards` 屬性設置為 `True`，使其沿著序列的相反方向運行。\n",
|
||||
"\n",
|
||||
"無論是單向還是雙向的循環網絡，都能捕捉序列中的模式，並將其存儲到狀態向量中或作為輸出返回。與卷積網絡類似，我們可以在第一個循環層之後再構建另一個循環層，以捕捉更高層次的模式，這些模式是由第一層提取的低層次模式構建而成的。這引出了 **多層 RNN** 的概念，它由兩層或更多層循環網絡組成，其中前一層的輸出作為下一層的輸入。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"*圖片來源於 [這篇精彩的文章](https://towardsdatascience.com/from-a-lstm-cell-to-a-multilayer-lstm-network-with-pytorch-2899eb5696f3)，作者是 Fernando López。*\n",
|
||||
"\n",
|
||||
"Keras 讓構建這些網絡變得非常簡單，因為你只需要在模型中添加更多的循環層。對於最後一層以外的所有層，我們需要指定 `return_sequences=True` 參數，因為我們需要該層返回所有的中間狀態，而不僅僅是循環計算的最終狀態。\n",
|
||||
"\n",
|
||||
"現在，我們來為分類問題構建一個雙層雙向 LSTM。\n",
|
||||
"\n",
|
||||
"> **Note** 這段代碼執行起來可能需要較長的時間，但它能給我們帶來目前為止最高的準確率。所以也許值得等待並看看結果。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 9,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"5044/7500 [===================>..........] - ETA: 2:33 - loss: 0.3709 - acc: 0.8706\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\r5045/7500 [===================>..........] - ETA: 2:33 - loss: 0.3709 - acc: 0.8706"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model = keras.models.Sequential([\n",
|
||||
" vectorizer,\n",
|
||||
" keras.layers.Embedding(vocab_size, 128, mask_zero=True),\n",
|
||||
" keras.layers.Bidirectional(keras.layers.LSTM(64,return_sequences=True)),\n",
|
||||
" keras.layers.Bidirectional(keras.layers.LSTM(64)), \n",
|
||||
" keras.layers.Dense(4,activation='softmax')\n",
|
||||
"])\n",
|
||||
"\n",
|
||||
"model.compile(loss='sparse_categorical_crossentropy',metrics=['acc'], optimizer='adam')\n",
|
||||
"model.fit(ds_train.map(tupelize).batch(batch_size),\n",
|
||||
" validation_data=ds_test.map(tupelize).batch(batch_size))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## RNNs 用於其他任務\n",
|
||||
"\n",
|
||||
"到目前為止，我們主要集中在使用 RNNs 來對文本序列進行分類。但它們還能處理更多任務，例如文本生成和機器翻譯——我們會在下一單元探討這些任務。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件已使用人工智能翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。應以原文文件作為權威來源。對於關鍵資訊，建議尋求專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或誤釋不承擔責任。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"kernel_info": {
|
||||
"name": "conda-env-py37_tensorflow-py"
|
||||
},
|
||||
"kernelspec": {
|
||||
"display_name": "py37_tensorflow",
|
||||
"language": "python",
|
||||
"name": "conda-env-py37_tensorflow-py"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.7.9"
|
||||
},
|
||||
"nteract": {
|
||||
"version": "nteract-front-end@1.0.0"
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "81351e61f619b432ff51010a4f993194",
|
||||
"translation_date": "2025-08-31T10:48:08+00:00",
|
||||
"source_file": "lessons/5-NLP/16-RNN/RNNTF.ipynb",
|
||||
"language_code": "hk"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 4
|
||||
}
|
||||
|
|
@ -0,0 +1,414 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 生成式網絡\n",
|
||||
"\n",
|
||||
"循環神經網絡（Recurrent Neural Networks, RNNs）及其門控單元變體，例如長短期記憶單元（Long Short Term Memory Cells, LSTMs）和門控循環單元（Gated Recurrent Units, GRUs），提供了一種語言建模的機制，也就是說，它們可以學習單詞的排列順序，並對序列中的下一個單詞進行預測。這使得我們可以使用 RNNs 來完成**生成任務**，例如普通文本生成、機器翻譯，甚至是圖像描述生成。\n",
|
||||
"\n",
|
||||
"在上一單元中討論的 RNN 結構中，每個 RNN 單元會輸出下一個隱藏狀態。然而，我們也可以為每個循環單元添加另一個輸出，這樣就可以輸出一個**序列**（其長度與原始序列相等）。此外，我們還可以使用不在每一步接受輸入的 RNN 單元，而是僅接受一個初始狀態向量，然後生成一個輸出序列。\n",
|
||||
"\n",
|
||||
"在這個筆記本中，我們將專注於幫助我們生成文本的簡單生成模型。為了簡化，我們將構建一個**字符級網絡**，逐字母生成文本。在訓練過程中，我們需要使用一些文本語料庫，並將其拆分為字母序列。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 1,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Loading dataset...\n",
|
||||
"Building vocab...\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"import torch\n",
|
||||
"import torchtext\n",
|
||||
"import numpy as np\n",
|
||||
"from torchnlp import *\n",
|
||||
"train_dataset,test_dataset,classes,vocab = load_dataset()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 建立字符詞彙表\n",
|
||||
"\n",
|
||||
"為了建立基於字符層級的生成網絡，我們需要將文本拆分為單個字符，而不是單詞。這可以通過定義一個不同的分詞器來完成：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 2,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Vocabulary size = 82\n",
|
||||
"Encoding of 'a' is 1\n",
|
||||
"Character with code 13 is c\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"def char_tokenizer(words):\n",
|
||||
" return list(words) #[word for word in words]\n",
|
||||
"\n",
|
||||
"counter = collections.Counter()\n",
|
||||
"for (label, line) in train_dataset:\n",
|
||||
" counter.update(char_tokenizer(line))\n",
|
||||
"vocab = torchtext.vocab.vocab(counter)\n",
|
||||
"\n",
|
||||
"vocab_size = len(vocab)\n",
|
||||
"print(f\"Vocabulary size = {vocab_size}\")\n",
|
||||
"print(f\"Encoding of 'a' is {vocab.get_stoi()['a']}\")\n",
|
||||
"print(f\"Character with code 13 is {vocab.get_itos()[13]}\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"讓我們看看如何對我們數據集中的文本進行編碼的例子：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"tensor([ 0, 1, 2, 2, 3, 4, 5, 6, 3, 7, 8, 1, 9, 10, 3, 11, 2, 1,\n",
|
||||
" 12, 3, 7, 1, 13, 14, 3, 15, 16, 5, 17, 3, 5, 18, 8, 3, 7, 2,\n",
|
||||
" 1, 13, 14, 3, 19, 20, 8, 21, 5, 8, 9, 10, 22, 3, 20, 8, 21, 5,\n",
|
||||
" 8, 9, 10, 3, 23, 3, 4, 18, 17, 9, 5, 23, 10, 8, 2, 2, 8, 9,\n",
|
||||
" 10, 24, 3, 0, 1, 2, 2, 3, 4, 5, 9, 8, 8, 5, 25, 10, 3, 26,\n",
|
||||
" 12, 27, 16, 26, 2, 27, 16, 28, 29, 30, 1, 16, 26, 3, 17, 31, 3, 21,\n",
|
||||
" 2, 5, 9, 1, 23, 13, 32, 16, 27, 13, 10, 24, 3, 1, 9, 8, 3, 10,\n",
|
||||
" 8, 8, 27, 16, 28, 3, 28, 9, 8, 8, 16, 3, 1, 28, 1, 27, 16, 6])"
|
||||
]
|
||||
},
|
||||
"execution_count": 3,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"def enc(x):\n",
|
||||
" return torch.LongTensor(encode(x,voc=vocab,tokenizer=char_tokenizer))\n",
|
||||
"\n",
|
||||
"enc(train_dataset[0][1])"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 訓練生成式 RNN\n",
|
||||
"\n",
|
||||
"我們訓練 RNN 生成文本的方法如下。在每一步中，我們會取一段長度為 `nchars` 的字符序列，並要求網絡為每個輸入字符生成下一個輸出字符：\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"根據實際情況，我們可能還需要加入一些特殊字符，例如 *序列結束符* `<eos>`。在我們的情況下，我們只希望訓練網絡進行無限文本生成，因此我們會將每段序列的大小固定為 `nchars` 個標記。因此，每個訓練樣本將包含 `nchars` 個輸入和 `nchars` 個輸出（輸入序列向左移動一個符號）。小批量訓練將由多個這樣的序列組成。\n",
|
||||
"\n",
|
||||
"我們生成小批量的方法是取每段長度為 `l` 的新聞文本，並從中生成所有可能的輸入-輸出組合（這樣的組合會有 `l-nchars` 個）。它們將形成一個小批量，而每次訓練步驟的小批量大小會有所不同。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"(tensor([[ 0, 1, 2, ..., 28, 29, 30],\n",
|
||||
" [ 1, 2, 2, ..., 29, 30, 1],\n",
|
||||
" [ 2, 2, 3, ..., 30, 1, 16],\n",
|
||||
" ...,\n",
|
||||
" [20, 8, 21, ..., 1, 28, 1],\n",
|
||||
" [ 8, 21, 5, ..., 28, 1, 27],\n",
|
||||
" [21, 5, 8, ..., 1, 27, 16]]),\n",
|
||||
" tensor([[ 1, 2, 2, ..., 29, 30, 1],\n",
|
||||
" [ 2, 2, 3, ..., 30, 1, 16],\n",
|
||||
" [ 2, 3, 4, ..., 1, 16, 26],\n",
|
||||
" ...,\n",
|
||||
" [ 8, 21, 5, ..., 28, 1, 27],\n",
|
||||
" [21, 5, 8, ..., 1, 27, 16],\n",
|
||||
" [ 5, 8, 9, ..., 27, 16, 6]]))"
|
||||
]
|
||||
},
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"nchars = 100\n",
|
||||
"\n",
|
||||
"def get_batch(s,nchars=nchars):\n",
|
||||
" ins = torch.zeros(len(s)-nchars,nchars,dtype=torch.long,device=device)\n",
|
||||
" outs = torch.zeros(len(s)-nchars,nchars,dtype=torch.long,device=device)\n",
|
||||
" for i in range(len(s)-nchars):\n",
|
||||
" ins[i] = enc(s[i:i+nchars])\n",
|
||||
" outs[i] = enc(s[i+1:i+nchars+1])\n",
|
||||
" return ins,outs\n",
|
||||
"\n",
|
||||
"get_batch(train_dataset[0][1])"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在讓我們定義生成器網絡。它可以基於我們在上一單元討論過的任何循環單元（簡單、LSTM 或 GRU）。在我們的例子中，我們將使用 LSTM。\n",
|
||||
"\n",
|
||||
"由於網絡以字符作為輸入，而詞彙量相對較小，我們不需要嵌入層，直接使用獨熱編碼的輸入即可進入 LSTM 單元。然而，因為我們以字符編號作為輸入，所以在傳遞給 LSTM 之前需要對它們進行獨熱編碼。這可以在 `forward` 傳遞過程中通過調用 `one_hot` 函數來完成。輸出編碼器將是一個線性層，用於將隱藏狀態轉換為獨熱編碼的輸出。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"class LSTMGenerator(torch.nn.Module):\n",
|
||||
" def __init__(self, vocab_size, hidden_dim):\n",
|
||||
" super().__init__()\n",
|
||||
" self.rnn = torch.nn.LSTM(vocab_size,hidden_dim,batch_first=True)\n",
|
||||
" self.fc = torch.nn.Linear(hidden_dim, vocab_size)\n",
|
||||
"\n",
|
||||
" def forward(self, x, s=None):\n",
|
||||
" x = torch.nn.functional.one_hot(x,vocab_size).to(torch.float32)\n",
|
||||
" x,s = self.rnn(x,s)\n",
|
||||
" return self.fc(x),s"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"在訓練過程中，我們希望能夠對生成的文本進行取樣。為此，我們將定義一個 `generate` 函數，該函數會生成一個長度為 `size` 的輸出字串，並以初始字串 `start` 開始。\n",
|
||||
"\n",
|
||||
"其運作方式如下：首先，我們會將整個初始字串傳遞到網絡中，並獲取輸出狀態 `s` 和下一個預測字符 `out`。由於 `out` 是以 one-hot 編碼表示的，我們使用 `argmax` 來獲取該字符在詞彙表中的索引 `nc`，然後使用 `itos` 找出實際字符，並將其附加到結果字符列表 `chars` 中。這個生成一個字符的過程會重複執行 `size` 次，以生成所需數量的字符。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 8,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"def generate(net,size=100,start='today '):\n",
|
||||
" chars = list(start)\n",
|
||||
" out, s = net(enc(chars).view(1,-1).to(device))\n",
|
||||
" for i in range(size):\n",
|
||||
" nc = torch.argmax(out[0][-1])\n",
|
||||
" chars.append(vocab.get_itos()[nc])\n",
|
||||
" out, s = net(nc.view(1,-1),s)\n",
|
||||
" return ''.join(chars)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在開始訓練吧！訓練循環幾乎與之前的例子相同，但這次我們每隔 1000 個 epoch 就會打印一次生成的文本，而不是準確率。\n",
|
||||
"\n",
|
||||
"需要特別注意的是計算損失的方式。我們需要根據 one-hot 編碼的輸出 `out` 和期望的文本 `text_out`（即字符索引的列表）來計算損失。幸運的是，`cross_entropy` 函數的第一個參數是未正規化的網絡輸出，第二個參數是類別編號，這正好符合我們的需求。它還會自動對小批量的大小進行平均。\n",
|
||||
"\n",
|
||||
"我們還通過 `samples_to_train` 限制訓練樣本的數量，以避免等待過久。我們鼓勵你進行實驗，嘗試更長時間的訓練，可能是多個 epoch（在這種情況下，你需要在這段代碼外再創建一個循環）。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 9,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Current loss = 4.398899078369141\n",
|
||||
"today sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr s\n",
|
||||
"Current loss = 2.161320447921753\n",
|
||||
"today and to the tor to to the tor to to the tor to to the tor to to the tor to to the tor to to the tor t\n",
|
||||
"Current loss = 1.6722588539123535\n",
|
||||
"today and the court to the could to the could to the could to the could to the could to the could to the c\n",
|
||||
"Current loss = 2.423795223236084\n",
|
||||
"today and a second to the conternation of the conternation of the conternation of the conternation of the \n",
|
||||
"Current loss = 1.702607274055481\n",
|
||||
"today and the company to the company to the company to the company to the company to the company to the co\n",
|
||||
"Current loss = 1.692358136177063\n",
|
||||
"today and the company to the company to the company to the company to the company to the company to the co\n",
|
||||
"Current loss = 1.9722288846969604\n",
|
||||
"today and the control the control the control the control the control the control the control the control \n",
|
||||
"Current loss = 1.8705692291259766\n",
|
||||
"today and the second to the second to the second to the second to the second to the second to the second t\n",
|
||||
"Current loss = 1.7626899480819702\n",
|
||||
"today and a security and a security and a security and a security and a security and a security and a secu\n",
|
||||
"Current loss = 1.5574463605880737\n",
|
||||
"today and the company and the company and the company and the company and the company and the company and \n",
|
||||
"Current loss = 1.5620026588439941\n",
|
||||
"today and the be that the be the be that the be the be that the be the be that the be the be that the be t\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"net = LSTMGenerator(vocab_size,64).to(device)\n",
|
||||
"\n",
|
||||
"samples_to_train = 10000\n",
|
||||
"optimizer = torch.optim.Adam(net.parameters(),0.01)\n",
|
||||
"loss_fn = torch.nn.CrossEntropyLoss()\n",
|
||||
"net.train()\n",
|
||||
"for i,x in enumerate(train_dataset):\n",
|
||||
" # x[0] is class label, x[1] is text\n",
|
||||
" if len(x[1])-nchars<10:\n",
|
||||
" continue\n",
|
||||
" samples_to_train-=1\n",
|
||||
" if not samples_to_train: break\n",
|
||||
" text_in, text_out = get_batch(x[1])\n",
|
||||
" optimizer.zero_grad()\n",
|
||||
" out,s = net(text_in)\n",
|
||||
" loss = torch.nn.functional.cross_entropy(out.view(-1,vocab_size),text_out.flatten()) #cross_entropy(out,labels)\n",
|
||||
" loss.backward()\n",
|
||||
" optimizer.step()\n",
|
||||
" if i%1000==0:\n",
|
||||
" print(f\"Current loss = {loss.item()}\")\n",
|
||||
" print(generate(net))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"這個例子已經生成了一些相當不錯的文本，但仍有幾個方面可以進一步改進：\n",
|
||||
"\n",
|
||||
"* **更好的小批次生成**。我們用來準備訓練數據的方法是從一個樣本中生成一個小批次。這並不理想，因為小批次的大小各不相同，有些甚至無法生成，因為文本小於 `nchars`。此外，過小的小批次無法充分利用 GPU 的性能。更明智的做法是從所有樣本中提取一大塊文本，然後生成所有的輸入-輸出對，將它們打亂，並生成大小相等的小批次。\n",
|
||||
"\n",
|
||||
"* **多層 LSTM**。嘗試使用 2 或 3 層的 LSTM 單元是有意義的。如我們在上一單元提到的，每一層 LSTM 都會從文本中提取某些模式，而在字符級生成器的情況下，我們可以預期較低層的 LSTM 負責提取音節，而較高層則負責單詞和單詞組合。這可以通過向 LSTM 構造函數傳遞層數參數來簡單實現。\n",
|
||||
"\n",
|
||||
"* 你也可以嘗試使用 **GRU 單元**，看看哪種表現更好，還可以嘗試 **不同的隱藏層大小**。過大的隱藏層可能導致過擬合（例如，網絡會學習到精確的文本），而過小的隱藏層可能無法產生良好的結果。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 軟文本生成與溫度\n",
|
||||
"\n",
|
||||
"在之前 `generate` 的定義中，我們總是選擇機率最高的字符作為生成文本中的下一個字符。這導致生成的文本經常在相同的字符序列之間「循環」出現，如以下例子所示：\n",
|
||||
"```\n",
|
||||
"today of the second the company and a second the company ...\n",
|
||||
"```\n",
|
||||
"\n",
|
||||
"然而，如果我們查看下一個字符的機率分佈，可能會發現幾個最高機率之間的差距並不大，例如一個字符的機率是 0.2，另一個是 0.19，等等。例如，在尋找序列 '*play*' 的下一個字符時，下一個字符可能同樣有可能是空格，或者是 **e**（如單詞 *player* 中的情況）。\n",
|
||||
"\n",
|
||||
"這讓我們得出一個結論：選擇機率最高的字符並不總是「公平」的，因為選擇第二高的字符仍然可能生成有意義的文本。更明智的做法是從網絡輸出的機率分佈中**抽樣**字符。\n",
|
||||
"\n",
|
||||
"這種抽樣可以使用 `multinomial` 函數來完成，該函數實現了所謂的**多項分佈**。下面定義了一個實現這種**軟**文本生成的函數：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 10,
|
||||
"metadata": {
|
||||
"scrolled": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"--- Temperature = 0.3\n",
|
||||
"Today and a company and complete an all the land the restrational the as a security and has provers the pay to and a report and the computer in the stand has filities and working the law the stations for a company and with the company and the final the first company and refight of the state and and workin\n",
|
||||
"\n",
|
||||
"--- Temperature = 0.8\n",
|
||||
"Today he oniis its first to Aus bomblaties the marmation a to manan boogot that pirate assaid a relaid their that goverfin the the Cappets Ecrotional Assonia Cition targets it annight the w scyments Blamity #39;s TVeer Diercheg Reserals fran envyuil that of ster said access what succers of Dour-provelith\n",
|
||||
"\n",
|
||||
"--- Temperature = 1.0\n",
|
||||
"Today holy they a 11 will meda a toket subsuaties, engins for Chanos, they's has stainger past to opening orital his thempting new Nattona was al innerforder advan-than #36;s night year his religuled talitatian what the but with Wednesday to Justment will wemen of Mark CCC Camp as Timed Nae wome a leaders\n",
|
||||
"\n",
|
||||
"--- Temperature = 1.3\n",
|
||||
"Today gpone 2.5 fech atcusion poor cocles toparsdorM.cht Line Pamage put 43 his calt lowed to the book, that has authh-the silia rruch ailing to'ory andhes beutirsimi- Aefffive heading offil an auf eacklets is charged evis, Gunymy oy) Mony has it after-sloythyor loveId out filme, the Natabl -Najuntaxiggs \n",
|
||||
"\n",
|
||||
"--- Temperature = 1.8\n",
|
||||
"Today plary, P.slan chly\\401 mardregationly #39;t 8.1Mide) closes ,filtcon alfly playin roven!\\grea.-QFBEP: Iss onfarchQ/itilia CCf Zivesigntwasta orce.-Peul-aw.uicrin of fuglinfsut aftaningwo, MIEX awayew Aice Woiduar Corvagiugge oppo esig ThusBratourid canthly-RyI.co lagitems\\eexciaishes.conBabntusmor I\n",
|
||||
"\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"def generate_soft(net,size=100,start='today ',temperature=1.0):\n",
|
||||
" chars = list(start)\n",
|
||||
" out, s = net(enc(chars).view(1,-1).to(device))\n",
|
||||
" for i in range(size):\n",
|
||||
" #nc = torch.argmax(out[0][-1])\n",
|
||||
" out_dist = out[0][-1].div(temperature).exp()\n",
|
||||
" nc = torch.multinomial(out_dist,1)[0]\n",
|
||||
" chars.append(vocab.get_itos()[nc])\n",
|
||||
" out, s = net(nc.view(1,-1),s)\n",
|
||||
" return ''.join(chars)\n",
|
||||
" \n",
|
||||
"for i in [0.3,0.8,1.0,1.3,1.8]:\n",
|
||||
" print(f\"--- Temperature = {i}\\n{generate_soft(net,size=300,start='Today ',temperature=i)}\\n\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們引入了一個名為 **temperature** 的參數，用於指示我們應該多大程度地堅持最高概率。如果 temperature 是 1.0，我們進行公平的多項式抽樣，而當 temperature 趨於無窮大時，所有概率變得相等，我們隨機選擇下一個字符。在下面的例子中，我們可以觀察到，當我們將 temperature 增加得太高時，文本變得毫無意義，而當它接近 0 時，則類似於「循環」的硬生成文本。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件已使用人工智能翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。應以原文文件作為權威來源。對於關鍵資訊，建議尋求專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或誤釋不承擔責任。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"interpreter": {
|
||||
"hash": "16af2a8bbb083ea23e5e41c7f5787656b2ce26968575d8763f2c4b17f9cd711f"
|
||||
},
|
||||
"kernelspec": {
|
||||
"display_name": "Python 3.8.12 ('py38')",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.8.12"
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "7673cd150d96c74c6d6011460094efb4",
|
||||
"translation_date": "2025-08-31T10:35:18+00:00",
|
||||
"source_file": "lessons/5-NLP/17-GenerativeNetworks/GenerativePyTorch.ipynb",
|
||||
"language_code": "hk"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 4
|
||||
}
|
||||
|
|
@ -0,0 +1,497 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 生成式網絡\n",
|
||||
"\n",
|
||||
"循環神經網絡（Recurrent Neural Networks, RNNs）及其門控單元變體，例如長短期記憶單元（Long Short Term Memory Cells, LSTMs）和門控循環單元（Gated Recurrent Units, GRUs），提供了一種語言建模的機制，也就是說，它們可以學習單詞的排列順序，並對序列中的下一個單詞進行預測。這使得我們可以使用 RNNs 來完成**生成任務**，例如普通文本生成、機器翻譯，甚至是圖像描述生成。\n",
|
||||
"\n",
|
||||
"在上一單元中討論的 RNN 結構中，每個 RNN 單元會輸出下一個隱藏狀態。然而，我們也可以為每個循環單元添加另一個輸出，這樣就可以輸出一個**序列**（其長度與原始序列相等）。此外，我們還可以使用不在每一步接受輸入的 RNN 單元，而僅僅接受一個初始狀態向量，然後生成一個輸出序列。\n",
|
||||
"\n",
|
||||
"在這個筆記本中，我們將專注於幫助我們生成文本的簡單生成模型。為了簡化，我們將構建一個**字符級網絡**，逐字母生成文本。在訓練過程中，我們需要使用一些文本語料庫，並將其拆分為字母序列。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 1,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import tensorflow as tf\n",
|
||||
"from tensorflow import keras\n",
|
||||
"import tensorflow_datasets as tfds\n",
|
||||
"import numpy as np\n",
|
||||
"\n",
|
||||
"ds_train, ds_test = tfds.load('ag_news_subset').values()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 建立字符詞彙表\n",
|
||||
"\n",
|
||||
"要建立字符級生成網絡，我們需要將文本拆分為單個字符，而不是單詞。我們之前使用的 `TextVectorization` 層無法做到這一點，因此我們有以下兩個選擇：\n",
|
||||
"\n",
|
||||
"* 手動載入文本並自行進行分詞，如 [這個官方 Keras 範例](https://keras.io/examples/generative/lstm_character_level_text_generation/) 中所示\n",
|
||||
"* 使用 `Tokenizer` 類進行字符級分詞。\n",
|
||||
"\n",
|
||||
"我們會選擇第二個選項。`Tokenizer` 也可以用於分詞成單詞，因此可以輕鬆地從字符級分詞切換到單詞級分詞。\n",
|
||||
"\n",
|
||||
"要進行字符級分詞，我們需要傳入 `char_level=True` 參數：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 2,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"def extract_text(x):\n",
|
||||
" return x['title']+' '+x['description']\n",
|
||||
"\n",
|
||||
"def tupelize(x):\n",
|
||||
" return (extract_text(x),x['label'])\n",
|
||||
"\n",
|
||||
"tokenizer = keras.preprocessing.text.Tokenizer(char_level=True,lower=False)\n",
|
||||
"tokenizer.fit_on_texts([x['title'].numpy().decode('utf-8') for x in ds_train])"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們還希望使用一個特殊的標記來表示**序列結束**，我們將其稱為 `<eos>`。讓我們手動將其添加到詞彙表中：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"eos_token = len(tokenizer.word_index)+1\n",
|
||||
"tokenizer.word_index['<eos>'] = eos_token\n",
|
||||
"\n",
|
||||
"vocab_size = eos_token + 1"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在，要將文本編碼為數字序列，我們可以使用：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"[[48, 2, 10, 10, 5, 44, 1, 25, 5, 8, 10, 13, 78]]"
|
||||
]
|
||||
},
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"tokenizer.texts_to_sequences(['Hello, world!'])"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 訓練生成式 RNN 來生成標題\n",
|
||||
"\n",
|
||||
"我們將以以下方式訓練 RNN 來生成新聞標題。在每一步中，我們會取一個標題，將其輸入到 RNN 中，並對於每個輸入的字符，要求網絡生成下一個輸出的字符：\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"對於序列中的最後一個字符，我們會要求網絡生成 `<eos>` 標記。\n",
|
||||
"\n",
|
||||
"我們在這裡使用的生成式 RNN 的主要不同之處在於，我們會從 RNN 的每一步輸出中取結果，而不僅僅是從最後一個單元格中取結果。這可以通過為 RNN 單元格指定 `return_sequences` 參數來實現。\n",
|
||||
"\n",
|
||||
"因此，在訓練過程中，網絡的輸入將是一個特定長度的編碼字符序列，而輸出將是一個相同長度的序列，但向後偏移一個元素並以 `<eos>` 結尾。小批量數據將由若干這樣的序列組成，我們需要使用**填充**來對齊所有序列。\n",
|
||||
"\n",
|
||||
"現在讓我們創建一些函數來幫助我們轉換數據集。由於我們希望在小批量層級上填充序列，我們會先通過調用 `.batch()` 將數據集分批，然後使用 `map` 來進行轉換。因此，轉換函數將以整個小批量作為參數：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"def title_batch(x):\n",
|
||||
" x = [t.numpy().decode('utf-8') for t in x]\n",
|
||||
" z = tokenizer.texts_to_sequences(x)\n",
|
||||
" z = tf.keras.preprocessing.sequence.pad_sequences(z)\n",
|
||||
" return tf.one_hot(z,vocab_size), tf.one_hot(tf.concat([z[:,1:],tf.constant(eos_token,shape=(len(z),1))],axis=1),vocab_size)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"一些重要的事情我們在這裡做：\n",
|
||||
"* 我們首先從字串張量中提取實際文本\n",
|
||||
"* `text_to_sequences` 將字串列表轉換為整數張量列表\n",
|
||||
"* `pad_sequences` 將這些張量填充到它們的最大長度\n",
|
||||
"* 最後，我們對所有字符進行 one-hot 編碼，並進行移位和附加 `<eos>`。我們很快就會看到為什麼需要 one-hot 編碼的字符\n",
|
||||
"\n",
|
||||
"然而，這個函數是 **Pythonic** 的，也就是說它無法自動轉換為 Tensorflow 的計算圖。如果我們直接在 `Dataset.map` 函數中使用這個函數，會出現錯誤。我們需要使用 `py_function` 包裝器來封裝這個 Pythonic 調用：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 6,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"def title_batch_fn(x):\n",
|
||||
" x = x['title']\n",
|
||||
" a,b = tf.py_function(title_batch,inp=[x],Tout=(tf.float32,tf.float32))\n",
|
||||
" return a,b"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"> **注意**：區分 Pythonic 和 Tensorflow 的轉換函數可能看起來有點複雜，你可能會疑惑為什麼我們不在將數據集傳遞給 `fit` 之前，使用標準的 Python 函數進行轉換。雖然這確實是可行的，但使用 `Dataset.map` 有一個巨大的優勢，因為數據轉換管道是通過 Tensorflow 的計算圖執行的，這可以利用 GPU 的計算能力，並且減少了在 CPU 和 GPU 之間傳遞數據的需求。\n",
|
||||
"\n",
|
||||
"現在我們可以建立生成器網絡並開始訓練。它可以基於我們在上一單元討論過的任何循環單元（簡單的、LSTM 或 GRU）。在我們的例子中，我們將使用 LSTM。\n",
|
||||
"\n",
|
||||
"由於網絡以字符作為輸入，且詞彙表的大小相對較小，我們不需要嵌入層，直接使用 one-hot 編碼的輸入進入 LSTM 單元即可。輸出層將是一個 `Dense` 分類器，用於將 LSTM 的輸出轉換為 one-hot 編碼的標籤數字。\n",
|
||||
"\n",
|
||||
"此外，因為我們處理的是可變長度的序列，我們可以使用 `Masking` 層來創建一個掩碼，忽略字符串中填充的部分。這並不是絕對必要的，因為我們對 `<eos>` 標籤之後的內容並不特別感興趣，但我們會使用它來獲得一些使用這類型層的經驗。`input_shape` 將是 `(None, vocab_size)`，其中 `None` 表示可變長度的序列，而輸出形狀也是 `(None, vocab_size)`，正如你可以從 `summary` 中看到的那樣。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Model: \"sequential\"\n",
|
||||
"_________________________________________________________________\n",
|
||||
"Layer (type) Output Shape Param # \n",
|
||||
"=================================================================\n",
|
||||
"masking (Masking) (None, None, 84) 0 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"lstm (LSTM) (None, None, 128) 109056 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"dense (Dense) (None, None, 84) 10836 \n",
|
||||
"=================================================================\n",
|
||||
"Total params: 119,892\n",
|
||||
"Trainable params: 119,892\n",
|
||||
"Non-trainable params: 0\n",
|
||||
"_________________________________________________________________\n",
|
||||
"15000/15000 [==============================] - 229s 15ms/step - loss: 1.5385\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<tensorflow.python.keras.callbacks.History at 0x7fa40c1245e0>"
|
||||
]
|
||||
},
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model = keras.models.Sequential([\n",
|
||||
" keras.layers.Masking(input_shape=(None,vocab_size)),\n",
|
||||
" keras.layers.LSTM(128,return_sequences=True),\n",
|
||||
" keras.layers.Dense(vocab_size,activation='softmax')\n",
|
||||
"])\n",
|
||||
"\n",
|
||||
"model.summary()\n",
|
||||
"model.compile(loss='categorical_crossentropy')\n",
|
||||
"\n",
|
||||
"model.fit(ds_train.batch(8).map(title_batch_fn))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 生成輸出\n",
|
||||
"\n",
|
||||
"現在我們已經訓練了模型，接下來我們希望使用它來生成一些輸出。首先，我們需要一種方法來解碼由一系列標記數字表示的文本。為此，我們可以使用 `tokenizer.sequences_to_texts` 函數；然而，這個方法在字符級標記化時效果不佳。因此，我們將從 tokenizer 中提取標記字典（稱為 `word_index`），構建一個反向映射，並編寫我們自己的解碼函數：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 10,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"reverse_map = {val:key for key, val in tokenizer.word_index.items()}\n",
|
||||
"\n",
|
||||
"def decode(x):\n",
|
||||
" return ''.join([reverse_map[t] for t in x])"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在，我們開始進行生成。我們會從某個字串 `start` 開始，將其編碼成一個序列 `inp`，然後在每一步調用我們的網絡來推斷下一個字符。\n",
|
||||
"\n",
|
||||
"網絡的輸出 `out` 是一個包含 `vocab_size` 個元素的向量，代表每個標記的概率。我們可以通過使用 `argmax` 找到最有可能的標記編號。接著，我們將這個字符附加到已生成的標記列表中，並繼續生成過程。這個生成一個字符的過程會重複執行 `size` 次，以生成所需數量的字符。如果在過程中遇到 `eos_token`，我們會提前終止生成。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 12,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"'Today #39;s lead to strike for the strike for the strike for the strike (AFP)'"
|
||||
]
|
||||
},
|
||||
"execution_count": 12,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"def generate(model,size=100,start='Today '):\n",
|
||||
" inp = tokenizer.texts_to_sequences([start])[0]\n",
|
||||
" chars = inp\n",
|
||||
" for i in range(size):\n",
|
||||
" out = model(tf.expand_dims(tf.one_hot(inp,vocab_size),0))[0][-1]\n",
|
||||
" nc = tf.argmax(out)\n",
|
||||
" if nc==eos_token:\n",
|
||||
" break\n",
|
||||
" chars.append(nc.numpy())\n",
|
||||
" inp = inp+[nc]\n",
|
||||
" return decode(chars)\n",
|
||||
" \n",
|
||||
"generate(model)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 在訓練期間抽樣輸出\n",
|
||||
"\n",
|
||||
"由於我們沒有任何有用的指標，例如 *準確率*，唯一能判斷模型是否有所改善的方法就是在訓練期間**抽樣**生成的字串。為了實現這一點，我們會使用**回調函數**，即可以傳遞給 `fit` 函數的函數，並在訓練過程中定期被調用。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 13,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Epoch 1/3\n",
|
||||
"15000/15000 [==============================] - 226s 15ms/step - loss: 1.2703\n",
|
||||
"Today #39;s a lead in the company for the strike\n",
|
||||
"Epoch 2/3\n",
|
||||
"15000/15000 [==============================] - 227s 15ms/step - loss: 1.2057\n",
|
||||
"Today #39;s the Market Service on Security Start (AP)\n",
|
||||
"Epoch 3/3\n",
|
||||
"15000/15000 [==============================] - 226s 15ms/step - loss: 1.1752\n",
|
||||
"Today #39;s a line on the strike to start for the start\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<tensorflow.python.keras.callbacks.History at 0x7fa40c74e3d0>"
|
||||
]
|
||||
},
|
||||
"execution_count": 13,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"sampling_callback = keras.callbacks.LambdaCallback(\n",
|
||||
" on_epoch_end = lambda batch, logs: print(generate(model))\n",
|
||||
")\n",
|
||||
"\n",
|
||||
"model.fit(ds_train.batch(8).map(title_batch_fn),callbacks=[sampling_callback],epochs=3)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"這個例子已經生成了一些相當不錯的文本，但仍有幾個方面可以進一步改進：\n",
|
||||
"\n",
|
||||
"* **更多文本**。我們只使用了標題來進行任務，但你可能想嘗試使用完整的文本。請記住，RNN在處理長序列時表現並不太好，因此可以考慮將它們拆分成較短的句子，或者始終在固定的序列長度（例如，`num_chars`，假設為256）上進行訓練。你可以嘗試將上面的例子改造成這樣的架構，並參考[官方 Keras 教程](https://keras.io/examples/generative/lstm_character_level_text_generation/)作為靈感。\n",
|
||||
"\n",
|
||||
"* **多層 LSTM**。可以嘗試使用2層或3層的LSTM單元。如我們在前一單元提到的，每一層LSTM會從文本中提取特定的模式，而在字符級生成器的情況下，我們可以預期較低層的LSTM負責提取音節，而較高層則負責提取單詞和單詞組合。這可以通過向LSTM構造函數傳遞層數參數來簡單實現。\n",
|
||||
"\n",
|
||||
"* 你可能還想嘗試使用**GRU單元**，看看哪種表現更好，或者嘗試**不同的隱藏層大小**。隱藏層過大可能導致過擬合（例如，網絡會學習到精確的文本），而過小的大小可能無法產生良好的結果。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 軟性文本生成與溫度\n",
|
||||
"\n",
|
||||
"在之前 `generate` 的定義中，我們總是選擇機率最高的字符作為生成文本中的下一個字符。這導致文本經常在相同的字符序列之間「循環」出現，如以下例子所示：\n",
|
||||
"```\n",
|
||||
"today of the second the company and a second the company ...\n",
|
||||
"```\n",
|
||||
"\n",
|
||||
"然而，如果我們查看下一個字符的機率分佈，可能會發現幾個最高機率之間的差距並不大，例如一個字符的機率是 0.2，另一個是 0.19，等等。例如，在尋找序列 *play* 的下一個字符時，下一個字符可能同樣有可能是空格，或者是 **e**（如單詞 *player* 中的情況）。\n",
|
||||
"\n",
|
||||
"這讓我們得出一個結論：選擇機率最高的字符並不總是「公平」的，因為選擇第二高的字符仍然可能生成有意義的文本。更明智的做法是從網絡輸出的機率分佈中**抽樣**字符。\n",
|
||||
"\n",
|
||||
"這種抽樣可以使用 `np.multinomial` 函數來完成，該函數實現了所謂的**多項分佈**。下面定義了一個實現這種**軟性**文本生成的函數：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 33,
|
||||
"metadata": {
|
||||
"scrolled": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"\n",
|
||||
"--- Temperature = 0.3\n",
|
||||
"Today #39;s strike #39; to start at the store return\n",
|
||||
"On Sunday PO to Be Data Profit Up (Reuters)\n",
|
||||
"Moscow, SP wins straight to the Microsoft #39;s control of the space start\n",
|
||||
"President olding of the blast start for the strike to pay <b>...</b>\n",
|
||||
"Little red riding hood ficed to the spam countered in European <b>...</b>\n",
|
||||
"\n",
|
||||
"--- Temperature = 0.8\n",
|
||||
"Today countie strikes ryder missile faces food market blut\n",
|
||||
"On Sunday collores lose-toppy of sale of Bullment in <b>...</b>\n",
|
||||
"Moscow, IBM Diffeiting in Afghan Software Hotels (Reuters)\n",
|
||||
"President Ol Luster for Profit Peaced Raised (AP)\n",
|
||||
"Little red riding hood dace on depart talks #39; bank up\n",
|
||||
"\n",
|
||||
"--- Temperature = 1.0\n",
|
||||
"Today wits House buiting debate fixes #39; supervice stake again\n",
|
||||
"On Sunday arling digital poaching In for level\n",
|
||||
"Moscow, DS Up 7, Top Proble Protest Caprey Mamarian Strike\n",
|
||||
"President teps help of roubler stepted lessabul-Dhalitics (AFP)\n",
|
||||
"Little red riding hood signs on cash in Carter-youb\n",
|
||||
"\n",
|
||||
"--- Temperature = 1.3\n",
|
||||
"Today wits flawer ro, pSIA figat's co DroftwavesIs Talo up\n",
|
||||
"On Sunday hround elitwing wint EU Powerburlinetien\n",
|
||||
"Moscow, Bazz #39;s sentries olymen winnelds' next for Olympite Huc?\n",
|
||||
"President lost securitys from power Elections in Smiltrials\n",
|
||||
"Little red riding hood vides profit, exponituity, profitmainalist-at said listers\n",
|
||||
"\n",
|
||||
"--- Temperature = 1.8\n",
|
||||
"Today #39;It: He deat: N.KA Asside\n",
|
||||
"On Sunday i arry Par aldeup patient Wo stele1\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"ename": "KeyError",
|
||||
"evalue": "0",
|
||||
"output_type": "error",
|
||||
"traceback": [
|
||||
"\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
|
||||
"\u001b[0;31mKeyError\u001b[0m Traceback (most recent call last)",
|
||||
"\u001b[0;32m<ipython-input-33-db32367a0feb>\u001b[0m in \u001b[0;36m<module>\u001b[0;34m\u001b[0m\n\u001b[1;32m 18\u001b[0m \u001b[0mprint\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34mf\"\\n--- Temperature = {i}\"\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 19\u001b[0m \u001b[0;32mfor\u001b[0m \u001b[0mj\u001b[0m \u001b[0;32min\u001b[0m \u001b[0mrange\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;36m5\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 20\u001b[0;31m \u001b[0mprint\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mgenerate_soft\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mmodel\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0msize\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0;36m300\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0mstart\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0mwords\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0mj\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0mtemperature\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0mi\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m",
|
||||
"\u001b[0;32m<ipython-input-33-db32367a0feb>\u001b[0m in \u001b[0;36mgenerate_soft\u001b[0;34m(model, size, start, temperature)\u001b[0m\n\u001b[1;32m 11\u001b[0m \u001b[0mchars\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mappend\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mnc\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 12\u001b[0m \u001b[0minp\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0minp\u001b[0m\u001b[0;34m+\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0mnc\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 13\u001b[0;31m \u001b[0;32mreturn\u001b[0m \u001b[0mdecode\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mchars\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m 14\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 15\u001b[0m \u001b[0mwords\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0;34m[\u001b[0m\u001b[0;34m'Today '\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m'On Sunday '\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m'Moscow, '\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m'President '\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m'Little red riding hood '\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
|
||||
"\u001b[0;32m<ipython-input-10-3f5fa6130b1d>\u001b[0m in \u001b[0;36mdecode\u001b[0;34m(x)\u001b[0m\n\u001b[1;32m 2\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 3\u001b[0m \u001b[0;32mdef\u001b[0m \u001b[0mdecode\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mx\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m----> 4\u001b[0;31m \u001b[0;32mreturn\u001b[0m \u001b[0;34m''\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mjoin\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0mreverse_map\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0mt\u001b[0m\u001b[0;34m]\u001b[0m \u001b[0;32mfor\u001b[0m \u001b[0mt\u001b[0m \u001b[0;32min\u001b[0m \u001b[0mx\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m",
|
||||
"\u001b[0;32m<ipython-input-10-3f5fa6130b1d>\u001b[0m in \u001b[0;36m<listcomp>\u001b[0;34m(.0)\u001b[0m\n\u001b[1;32m 2\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 3\u001b[0m \u001b[0;32mdef\u001b[0m \u001b[0mdecode\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mx\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m----> 4\u001b[0;31m \u001b[0;32mreturn\u001b[0m \u001b[0;34m''\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mjoin\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0mreverse_map\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0mt\u001b[0m\u001b[0;34m]\u001b[0m \u001b[0;32mfor\u001b[0m \u001b[0mt\u001b[0m \u001b[0;32min\u001b[0m \u001b[0mx\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m",
|
||||
"\u001b[0;31mKeyError\u001b[0m: 0"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"def generate_soft(model,size=100,start='Today ',temperature=1.0):\n",
|
||||
" inp = tokenizer.texts_to_sequences([start])[0]\n",
|
||||
" chars = inp\n",
|
||||
" for i in range(size):\n",
|
||||
" out = model(tf.expand_dims(tf.one_hot(inp,vocab_size),0))[0][-1]\n",
|
||||
" probs = tf.exp(tf.math.log(out)/temperature).numpy().astype(np.float64)\n",
|
||||
" probs = probs/np.sum(probs)\n",
|
||||
" nc = np.argmax(np.random.multinomial(1,probs,1))\n",
|
||||
" if nc==eos_token:\n",
|
||||
" break\n",
|
||||
" chars.append(nc)\n",
|
||||
" inp = inp+[nc]\n",
|
||||
" return decode(chars)\n",
|
||||
"\n",
|
||||
"words = ['Today ','On Sunday ','Moscow, ','President ','Little red riding hood ']\n",
|
||||
" \n",
|
||||
"for i in [0.3,0.8,1.0,1.3,1.8]:\n",
|
||||
" print(f\"\\n--- Temperature = {i}\")\n",
|
||||
" for j in range(5):\n",
|
||||
" print(generate_soft(model,size=300,start=words[j],temperature=i))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們引入了一個名為 **temperature** 的參數，用於指示我們應該多大程度地堅持最高概率。如果 temperature 是 1.0，我們進行公平的多項式抽樣，而當 temperature 趨於無窮大時，所有概率變得相等，我們隨機選擇下一個字符。在下面的例子中，我們可以觀察到當我們將 temperature 增加得太多時，文本變得毫無意義，而當它接近 0 時，則類似於「循環」的硬生成文本。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件已使用人工智能翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。儘管我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。原始語言的文件應被視為權威來源。對於重要資訊，建議使用專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或錯誤解釋概不負責。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"interpreter": {
|
||||
"hash": "16af2a8bbb083ea23e5e41c7f5787656b2ce26968575d8763f2c4b17f9cd711f"
|
||||
},
|
||||
"kernelspec": {
|
||||
"display_name": "Python 3.8.12 ('py38')",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.8.12"
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "9fbb7d5fda708537649f71f5f646fcde",
|
||||
"translation_date": "2025-08-31T10:33:22+00:00",
|
||||
"source_file": "lessons/5-NLP/17-GenerativeNetworks/GenerativeTF.ipynb",
|
||||
"language_code": "hk"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 4
|
||||
}
|
||||
|
|
@ -0,0 +1,353 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 注意力機制與Transformer\n",
|
||||
"\n",
|
||||
"循環神經網絡的一個主要缺點是序列中的所有詞語對結果的影響相同。這導致標準的LSTM編碼器-解碼器模型在序列到序列任務（例如命名實體識別和機器翻譯）中的表現不佳。實際上，輸入序列中的某些特定詞語通常比其他詞語對序列輸出有更大的影響。\n",
|
||||
"\n",
|
||||
"考慮一個序列到序列模型，例如機器翻譯。它由兩個循環神經網絡實現，其中一個網絡（**編碼器**）將輸入序列壓縮成隱藏狀態，另一個網絡（**解碼器**）將隱藏狀態展開成翻譯結果。這種方法的問題在於，網絡的最終狀態很難記住句子的開頭部分，從而導致模型在處理長句子時質量下降。\n",
|
||||
"\n",
|
||||
"**注意力機制**提供了一種方法，能夠對每個輸入向量對RNN每個輸出預測的上下文影響進行加權。其實現方式是通過在輸入RNN的中間狀態和輸出RNN之間創建捷徑。以此方式，在生成輸出符號$y_t$時，我們會考慮所有輸入隱藏狀態$h_i$，並賦予不同的權重係數$\\alpha_{t,i}$。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"*具有加性注意力機制的編碼器-解碼器模型，來自 [Bahdanau et al., 2015](https://arxiv.org/pdf/1409.0473.pdf)，引用自 [這篇博客文章](https://lilianweng.github.io/lil-log/2018/06/24/attention-attention.html)*\n",
|
||||
"\n",
|
||||
"注意力矩陣$\\{\\alpha_{i,j}\\}$表示某些輸入詞語在生成輸出序列中的某個詞語時所起的作用程度。以下是這樣一個矩陣的示例：\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"*圖片取自 [Bahdanau et al., 2015](https://arxiv.org/pdf/1409.0473.pdf) (圖3)*\n",
|
||||
"\n",
|
||||
"注意力機制是目前或接近目前自然語言處理領域的最先進技術的核心。添加注意力機制會大幅增加模型參數的數量，這導致了RNN的擴展問題。RNN擴展的一個關鍵限制是模型的循環性使得批量處理和並行化訓練變得困難。在RNN中，序列的每個元素都需要按順序處理，這意味著它無法輕易並行化。\n",
|
||||
"\n",
|
||||
"注意力機制的採用結合這一限制，促成了如今我們所熟知和使用的最先進Transformer模型的誕生，從BERT到OpenGPT3。\n",
|
||||
"\n",
|
||||
"## Transformer模型\n",
|
||||
"\n",
|
||||
"與將每次預測的上下文傳遞到下一個評估步驟不同，**Transformer模型**使用**位置編碼**和注意力來捕捉給定輸入在提供的文本窗口內的上下文。下圖展示了位置編碼與注意力如何在給定窗口內捕捉上下文。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"由於每個輸入位置獨立映射到每個輸出位置，Transformer比RNN更容易並行化，這使得能夠構建更大、更具表達力的語言模型。每個注意力頭可以用來學習詞語之間的不同關係，從而改善下游的自然語言處理任務。\n",
|
||||
"\n",
|
||||
"**BERT**（Bidirectional Encoder Representations from Transformers，雙向編碼器表示）是一個非常大的多層Transformer網絡，*BERT-base*有12層，*BERT-large*有24層。該模型首先在大規模文本數據（維基百科+書籍）上進行無監督訓練（預測句子中的被遮蔽詞語）。在預訓練過程中，模型吸收了大量的語言理解能力，隨後可以通過微調其他數據集來利用這些能力。這個過程稱為**遷移學習**。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"Transformer架構有許多變體，包括BERT、DistilBERT、BigBird、OpenGPT3等，這些模型都可以進行微調。[HuggingFace套件](https://github.com/huggingface/)提供了使用PyTorch訓練這些架構的資源庫。\n",
|
||||
"\n",
|
||||
"## 使用BERT進行文本分類\n",
|
||||
"\n",
|
||||
"讓我們看看如何使用預訓練的BERT模型來解決我們的傳統任務：序列分類。我們將對原始的AG News數據集進行分類。\n",
|
||||
"\n",
|
||||
"首先，載入HuggingFace庫和我們的數據集：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 10,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Loading dataset...\n",
|
||||
"Building vocab...\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"import torch\n",
|
||||
"import torchtext\n",
|
||||
"from torchnlp import *\n",
|
||||
"import transformers\n",
|
||||
"train_dataset, test_dataset, classes, vocab = load_dataset()\n",
|
||||
"vocab_len = len(vocab)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"因為我們將使用預訓練的 BERT 模型，所以需要使用特定的 tokenizer。首先，我們會載入與預訓練 BERT 模型相關聯的 tokenizer。\n",
|
||||
"\n",
|
||||
"HuggingFace 庫包含一個預訓練模型的資源庫，你只需在 `from_pretrained` 函數中指定模型名稱作為參數即可使用。所有模型所需的二進制文件會自動下載。\n",
|
||||
"\n",
|
||||
"然而，有時你可能需要載入自己的模型，在這種情況下，你可以指定包含所有相關文件的目錄，包括 tokenizer 的參數、包含模型參數的 `config.json` 文件、二進制權重等。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 11,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"# To load the model from Internet repository using model name. \n",
|
||||
"# Use this if you are running from your own copy of the notebooks\n",
|
||||
"bert_model = 'bert-base-uncased' \n",
|
||||
"\n",
|
||||
"# To load the model from the directory on disk. Use this for Microsoft Learn module, because we have\n",
|
||||
"# prepared all required files for you.\n",
|
||||
"bert_model = './bert'\n",
|
||||
"\n",
|
||||
"tokenizer = transformers.BertTokenizer.from_pretrained(bert_model)\n",
|
||||
"\n",
|
||||
"MAX_SEQ_LEN = 128\n",
|
||||
"PAD_INDEX = tokenizer.convert_tokens_to_ids(tokenizer.pad_token)\n",
|
||||
"UNK_INDEX = tokenizer.convert_tokens_to_ids(tokenizer.unk_token)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"`tokenizer` 對象包含 `encode` 函數，可直接用於編碼文本：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 15,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"[101, 1052, 22123, 2953, 2818, 2003, 1037, 2307, 7705, 2005, 17953, 2361, 102]"
|
||||
]
|
||||
},
|
||||
"execution_count": 15,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"tokenizer.encode('PyTorch is a great framework for NLP')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"然後，我們來建立在訓練期間用於訪問數據的迭代器。由於 BERT 使用其自己的編碼函數，我們需要定義一個類似於之前定義的 `padify` 的填充函數：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"def pad_bert(b):\n",
|
||||
" # b is the list of tuples of length batch_size\n",
|
||||
" # - first element of a tuple = label, \n",
|
||||
" # - second = feature (text sequence)\n",
|
||||
" # build vectorized sequence\n",
|
||||
" v = [tokenizer.encode(x[1]) for x in b]\n",
|
||||
" # compute max length of a sequence in this minibatch\n",
|
||||
" l = max(map(len,v))\n",
|
||||
" return ( # tuple of two tensors - labels and features\n",
|
||||
" torch.LongTensor([t[0] for t in b]),\n",
|
||||
" torch.stack([torch.nn.functional.pad(torch.tensor(t),(0,l-len(t)),mode='constant',value=0) for t in v])\n",
|
||||
" )\n",
|
||||
"\n",
|
||||
"train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=8, collate_fn=pad_bert, shuffle=True)\n",
|
||||
"test_loader = torch.utils.data.DataLoader(test_dataset, batch_size=8, collate_fn=pad_bert)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"在我們的情況下，我們將使用名為 `bert-base-uncased` 的預訓練 BERT 模型。讓我們使用 `BertForSequenceClassfication` 套件來加載模型。這確保了我們的模型已經具備分類所需的架構，包括最終的分類器。您會看到一條警告訊息，指出最終分類器的權重尚未初始化，模型需要進行預訓練——這完全沒問題，因為這正是我們即將要做的！\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 9,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stderr",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Some weights of the model checkpoint at ./bert were not used when initializing BertForSequenceClassification: ['cls.predictions.bias', 'cls.predictions.transform.dense.weight', 'cls.predictions.transform.dense.bias', 'cls.predictions.decoder.weight', 'cls.seq_relationship.weight', 'cls.seq_relationship.bias', 'cls.predictions.transform.LayerNorm.weight', 'cls.predictions.transform.LayerNorm.bias']\n",
|
||||
"- This IS expected if you are initializing BertForSequenceClassification from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).\n",
|
||||
"- This IS NOT expected if you are initializing BertForSequenceClassification from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).\n",
|
||||
"Some weights of BertForSequenceClassification were not initialized from the model checkpoint at ./bert and are newly initialized: ['classifier.weight', 'classifier.bias']\n",
|
||||
"You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model = transformers.BertForSequenceClassification.from_pretrained(bert_model,num_labels=4).to(device)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在我們準備開始訓練了！由於 BERT 已經是預訓練模型，我們希望使用較小的學習率開始訓練，以免破壞初始權重。\n",
|
||||
"\n",
|
||||
"所有的繁重工作都由 `BertForSequenceClassification` 模型完成。當我們將訓練數據傳入模型時，它會返回損失值和輸入小批量數據的網絡輸出。我們使用損失值進行參數優化（`loss.backward()` 負責反向傳播），而使用 `out` 計算訓練準確率，方法是將獲得的標籤 `labs`（通過 `argmax` 計算）與預期的 `labels` 進行比較。\n",
|
||||
"\n",
|
||||
"為了控制訓練過程，我們會在多次迭代中累積損失值和準確率，並在每 `report_freq` 次訓練週期後打印出來。\n",
|
||||
"\n",
|
||||
"這次訓練可能需要相當長的時間，因此我們會限制迭代次數。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 6,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Loss = 1.1254194641113282, Accuracy = 0.585\n",
|
||||
"Loss = 0.6194715118408203, Accuracy = 0.83\n",
|
||||
"Loss = 0.46665248870849607, Accuracy = 0.8475\n",
|
||||
"Loss = 0.4309701919555664, Accuracy = 0.8575\n",
|
||||
"Loss = 0.35427074432373046, Accuracy = 0.8825\n",
|
||||
"Loss = 0.3306886291503906, Accuracy = 0.8975\n",
|
||||
"Loss = 0.30340143203735354, Accuracy = 0.8975\n",
|
||||
"Loss = 0.26139299392700194, Accuracy = 0.915\n",
|
||||
"Loss = 0.26708646774291994, Accuracy = 0.9225\n",
|
||||
"Loss = 0.3667240524291992, Accuracy = 0.8675\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"optimizer = torch.optim.Adam(model.parameters(), lr=2e-5)\n",
|
||||
"\n",
|
||||
"report_freq = 50\n",
|
||||
"iterations = 500 # make this larger to train for longer time!\n",
|
||||
"\n",
|
||||
"model.train()\n",
|
||||
"\n",
|
||||
"i,c = 0,0\n",
|
||||
"acc_loss = 0\n",
|
||||
"acc_acc = 0\n",
|
||||
"\n",
|
||||
"for labels,texts in train_loader:\n",
|
||||
" labels = labels.to(device)-1 # get labels in the range 0-3 \n",
|
||||
" texts = texts.to(device)\n",
|
||||
" loss, out = model(texts, labels=labels)[:2]\n",
|
||||
" labs = out.argmax(dim=1)\n",
|
||||
" acc = torch.mean((labs==labels).type(torch.float32))\n",
|
||||
" optimizer.zero_grad()\n",
|
||||
" loss.backward()\n",
|
||||
" optimizer.step()\n",
|
||||
" acc_loss += loss\n",
|
||||
" acc_acc += acc\n",
|
||||
" i+=1\n",
|
||||
" c+=1\n",
|
||||
" if i%report_freq==0:\n",
|
||||
" print(f\"Loss = {acc_loss.item()/c}, Accuracy = {acc_acc.item()/c}\")\n",
|
||||
" c = 0\n",
|
||||
" acc_loss = 0\n",
|
||||
" acc_acc = 0\n",
|
||||
" iterations-=1\n",
|
||||
" if not iterations:\n",
|
||||
" break"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"你可以看到（特別是當你增加迭代次數並等待足夠長的時間），BERT 分類器能夠提供相當不錯的準確度！這是因為 BERT 已經對語言的結構有相當深入的理解，我們只需要對最終的分類器進行微調即可。然而，由於 BERT 是一個大型模型，整個訓練過程需要花費很長的時間，並且需要強大的計算能力！（GPU，最好是多個 GPU）。\n",
|
||||
"\n",
|
||||
"> **注意：** 在我們的例子中，我們使用的是其中一個最小的預訓練 BERT 模型。還有更大的模型可能會帶來更好的結果。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 評估模型表現\n",
|
||||
"\n",
|
||||
"現在我們可以在測試數據集上評估模型的表現。評估的流程與訓練流程非常相似，但我們不能忘記透過呼叫 `model.eval()` 將模型切換到評估模式。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 10,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Final accuracy: 0.9047029702970297\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model.eval()\n",
|
||||
"iterations = 100\n",
|
||||
"acc = 0\n",
|
||||
"i = 0\n",
|
||||
"for labels,texts in test_loader:\n",
|
||||
" labels = labels.to(device)-1 \n",
|
||||
" texts = texts.to(device)\n",
|
||||
" _, out = model(texts, labels=labels)[:2]\n",
|
||||
" labs = out.argmax(dim=1)\n",
|
||||
" acc += torch.mean((labs==labels).type(torch.float32))\n",
|
||||
" i+=1\n",
|
||||
" if i>iterations: break\n",
|
||||
" \n",
|
||||
"print(f\"Final accuracy: {acc.item()/i}\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 重點\n",
|
||||
"\n",
|
||||
"在本單元中，我們已經看到如何輕鬆地從 **transformers** 庫中取用預訓練的語言模型，並將其調整至我們的文本分類任務。同樣地，BERT 模型也可以用於實體抽取、問題回答以及其他 NLP 任務。\n",
|
||||
"\n",
|
||||
"Transformer 模型代表了 NLP 領域的最新技術，在大多數情況下，當你開始嘗試實現自定義 NLP 解決方案時，它應該是首選。然而，如果你希望構建更高級的神經模型，理解本模組中討論的循環神經網絡的基本原理是非常重要的。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件已使用人工智能翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。雖然我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。原始語言的文件應被視為權威來源。對於重要資訊，建議使用專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或錯誤解釋概不負責。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"kernelspec": {
|
||||
"display_name": "py37_pytorch",
|
||||
"language": "python",
|
||||
"name": "conda-env-py37_pytorch-py"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.7.7"
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "753865967678a92dbce7d7efbd36d980",
|
||||
"translation_date": "2025-08-31T10:40:24+00:00",
|
||||
"source_file": "lessons/5-NLP/18-Transformers/TransformersPyTorch.ipynb",
|
||||
"language_code": "hk"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 4
|
||||
}
|
||||
|
|
@ -0,0 +1,819 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 注意力機制與Transformer模型\n",
|
||||
"\n",
|
||||
"循環神經網絡的一個主要缺點是序列中的所有詞語對結果的影響力相同。這導致標準的LSTM編碼器-解碼器模型在處理序列到序列任務（例如命名實體識別和機器翻譯）時表現不佳。實際上，輸入序列中的某些特定詞語通常比其他詞語對輸出序列的影響更大。\n",
|
||||
"\n",
|
||||
"以序列到序列模型（例如機器翻譯）為例。這種模型由兩個循環神經網絡實現，其中一個網絡（**編碼器**）將輸入序列壓縮成隱藏狀態，另一個網絡（**解碼器**）將隱藏狀態展開成翻譯結果。這種方法的問題在於，網絡的最終狀態很難記住句子的開頭部分，因此在處理長句子時模型的質量會下降。\n",
|
||||
"\n",
|
||||
"**注意力機制**提供了一種方法，能夠對每個輸入向量在RNN輸出預測中的上下文影響進行加權。其實現方式是通過在輸入RNN的中間狀態和輸出RNN之間創建捷徑。在生成輸出符號$y_t$時，我們會考慮所有輸入隱藏狀態$h_i$，並賦予不同的權重係數$\\alpha_{t,i}$。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"*帶有加性注意力機制的編碼器-解碼器模型，來自[Bahdanau et al., 2015](https://arxiv.org/pdf/1409.0473.pdf)，引用自[這篇博客文章](https://lilianweng.github.io/lil-log/2018/06/24/attention-attention.html)*\n",
|
||||
"\n",
|
||||
"注意力矩陣$\\{\\alpha_{i,j}\\}$表示某些輸入詞語在生成輸出序列中的某個詞語時所起的作用程度。以下是這樣一個矩陣的示例：\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"*圖片取自[Bahdanau et al., 2015](https://arxiv.org/pdf/1409.0473.pdf)（圖3）*\n",
|
||||
"\n",
|
||||
"注意力機制是目前或接近目前自然語言處理領域的最先進技術的核心。儘管添加注意力機制會大幅增加模型參數的數量，但這也導致了RNN的擴展問題。RNN擴展的一個關鍵限制是模型的循環特性使得批量處理和訓練並行化變得困難。在RNN中，序列的每個元素都需要按順序處理，這意味著它無法輕易並行化。\n",
|
||||
"\n",
|
||||
"注意力機制的採用結合了這一限制，促成了如今最先進的Transformer模型的誕生，這些模型包括BERT和OpenGPT3等。\n",
|
||||
"\n",
|
||||
"## Transformer模型\n",
|
||||
"\n",
|
||||
"與將每次預測的上下文傳遞到下一個評估步驟不同，**Transformer模型**使用**位置編碼**和**注意力**來捕捉給定輸入在指定文本窗口內的上下文。下圖展示了位置編碼與注意力如何在給定窗口內捕捉上下文。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"由於每個輸入位置可以獨立映射到每個輸出位置，Transformer模型比RNN更容易並行化，這使得能夠構建更大、更具表達力的語言模型。每個注意力頭可以用來學習詞語之間的不同關係，從而改善下游的自然語言處理任務。\n",
|
||||
"\n",
|
||||
"## 構建簡單的Transformer模型\n",
|
||||
"\n",
|
||||
"Keras並不包含內建的Transformer層，但我們可以自己構建。與之前一樣，我們將專注於AG News數據集的文本分類，但值得一提的是，Transformer模型在更困難的自然語言處理任務中表現最佳。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 1,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import tensorflow as tf\n",
|
||||
"from tensorflow import keras\n",
|
||||
"import tensorflow_datasets as tfds\n",
|
||||
"import numpy as np\n",
|
||||
"\n",
|
||||
"ds_train, ds_test = tfds.load('ag_news_subset').values()\n",
|
||||
"\n",
|
||||
"def extract_text(x):\n",
|
||||
" return x['title']+' '+x['description']\n",
|
||||
"\n",
|
||||
"def tupelize(x):\n",
|
||||
" return (extract_text(x),x['label'])"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"新的 Keras 層應該繼承 `Layer` 類別，並實現 `call` 方法。我們先從 **位置嵌入** 層開始。我們將使用[官方 Keras 文件中的一些代碼](https://keras.io/examples/nlp/text_classification_with_transformer/)。我們假設我們將所有輸入序列填充到長度 `maxlen`。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 2,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"class TokenAndPositionEmbedding(keras.layers.Layer):\n",
|
||||
" def __init__(self, maxlen, vocab_size, embed_dim):\n",
|
||||
" super(TokenAndPositionEmbedding, self).__init__()\n",
|
||||
" self.token_emb = keras.layers.Embedding(input_dim=vocab_size, output_dim=embed_dim)\n",
|
||||
" self.pos_emb = keras.layers.Embedding(input_dim=maxlen, output_dim=embed_dim)\n",
|
||||
" self.maxlen = maxlen\n",
|
||||
"\n",
|
||||
" def call(self, x):\n",
|
||||
" maxlen = self.maxlen\n",
|
||||
" positions = tf.range(start=0, limit=maxlen, delta=1)\n",
|
||||
" positions = self.pos_emb(positions)\n",
|
||||
" x = self.token_emb(x)\n",
|
||||
" return x+positions"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"這一層包含兩個 `Embedding` 層：一個用於嵌入標記（以我們之前討論過的方式），另一個用於嵌入標記的位置。標記位置是通過使用 `tf.range` 從 0 到 `maxlen` 創建的一個自然數序列，然後傳遞到嵌入層。兩個生成的嵌入向量隨後相加，產生輸入的帶位置嵌入的表示，其形狀為 `maxlen`$\\times$`embed_dim`。\n",
|
||||
"\n",
|
||||
"現在，我們來實現 transformer 區塊。它將接收之前定義的嵌入層的輸出：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"class TransformerBlock(keras.layers.Layer):\n",
|
||||
" def __init__(self, embed_dim, num_heads, ff_dim, rate=0.1):\n",
|
||||
" super(TransformerBlock, self).__init__()\n",
|
||||
" self.att = keras.layers.MultiHeadAttention(num_heads=num_heads, key_dim=embed_dim, name='attn')\n",
|
||||
" self.ffn = keras.Sequential(\n",
|
||||
" [keras.layers.Dense(ff_dim, activation=\"relu\"), keras.layers.Dense(embed_dim),]\n",
|
||||
" )\n",
|
||||
" self.layernorm1 = keras.layers.LayerNormalization(epsilon=1e-6)\n",
|
||||
" self.layernorm2 = keras.layers.LayerNormalization(epsilon=1e-6)\n",
|
||||
" self.dropout1 = keras.layers.Dropout(rate)\n",
|
||||
" self.dropout2 = keras.layers.Dropout(rate)\n",
|
||||
"\n",
|
||||
" def call(self, inputs, training):\n",
|
||||
" attn_output = self.att(inputs, inputs)\n",
|
||||
" attn_output = self.dropout1(attn_output, training=training)\n",
|
||||
" out1 = self.layernorm1(inputs + attn_output)\n",
|
||||
" ffn_output = self.ffn(out1)\n",
|
||||
" ffn_output = self.dropout2(ffn_output, training=training)\n",
|
||||
" return self.layernorm2(out1 + ffn_output)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在，我們準備好定義完整的 Transformer 模型：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Model: \"sequential_1\"\n",
|
||||
"_________________________________________________________________\n",
|
||||
"Layer (type) Output Shape Param # \n",
|
||||
"=================================================================\n",
|
||||
"text_vectorization (TextVect (None, 256) 0 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"token_and_position_embedding (None, 256, 32) 648192 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"transformer_block (Transform (None, 256, 32) 10656 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"global_average_pooling1d (Gl (None, 32) 0 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"dropout_2 (Dropout) (None, 32) 0 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"dense_2 (Dense) (None, 20) 660 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"dropout_3 (Dropout) (None, 20) 0 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"dense_3 (Dense) (None, 4) 84 \n",
|
||||
"=================================================================\n",
|
||||
"Total params: 659,592\n",
|
||||
"Trainable params: 659,592\n",
|
||||
"Non-trainable params: 0\n",
|
||||
"_________________________________________________________________\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"embed_dim = 32 # Embedding size for each token\n",
|
||||
"num_heads = 2 # Number of attention heads\n",
|
||||
"ff_dim = 32 # Hidden layer size in feed forward network inside transformer\n",
|
||||
"maxlen = 256\n",
|
||||
"vocab_size = 20000\n",
|
||||
"\n",
|
||||
"model = keras.models.Sequential([\n",
|
||||
" keras.layers.experimental.preprocessing.TextVectorization(max_tokens=vocab_size,output_sequence_length=maxlen, input_shape=(1,)),\n",
|
||||
" TokenAndPositionEmbedding(maxlen, vocab_size, embed_dim),\n",
|
||||
" TransformerBlock(embed_dim, num_heads, ff_dim),\n",
|
||||
" keras.layers.GlobalAveragePooling1D(),\n",
|
||||
" keras.layers.Dropout(0.1),\n",
|
||||
" keras.layers.Dense(20, activation=\"relu\"),\n",
|
||||
" keras.layers.Dropout(0.1),\n",
|
||||
" keras.layers.Dense(4, activation=\"softmax\")\n",
|
||||
"])\n",
|
||||
"\n",
|
||||
"model.summary()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Training tokenizer\n",
|
||||
"938/938 [==============================] - 45s 39ms/step - loss: 0.4978 - acc: 0.8068 - val_loss: 0.2808 - val_acc: 0.9124\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<tensorflow.python.keras.callbacks.History at 0x7f9c2427a0d0>"
|
||||
]
|
||||
},
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"print('Training tokenizer')\n",
|
||||
"model.layers[0].adapt(ds_train.map(extract_text))\n",
|
||||
"model.compile(loss='sparse_categorical_crossentropy',metrics=['acc'], optimizer='adam')\n",
|
||||
"model.fit(ds_train.map(tupelize).batch(128),validation_data=ds_test.map(tupelize).batch(128))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## BERT Transformer 模型\n",
|
||||
"\n",
|
||||
"**BERT**（Bidirectional Encoder Representations from Transformers，雙向編碼器表示）是一個非常大型的多層 Transformer 網絡，*BERT-base* 有 12 層，而 *BERT-large* 則有 24 層。該模型首先在大規模文本數據（維基百科 + 書籍）上進行無監督訓練（預測句子中被遮蔽的詞語）。在預訓練過程中，模型吸收了大量的語言理解能力，然後可以通過微調其他數據集來加以利用。這個過程被稱為 **遷移學習**。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"Transformer 架構有許多變體，包括 BERT、DistilBERT、BigBird、OpenGPT3 等，這些模型都可以進行微調。\n",
|
||||
"\n",
|
||||
"現在讓我們看看如何使用預訓練的 BERT 模型來解決我們傳統的序列分類問題。我們將借用[官方文檔](https://www.tensorflow.org/text/tutorials/classify_text_with_bert)中的一些想法和代碼。\n",
|
||||
"\n",
|
||||
"為了加載預訓練模型，我們將使用 **Tensorflow hub**。首先，讓我們加載 BERT 專用的向量化工具：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 1,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"ename": "ModuleNotFoundError",
|
||||
"evalue": "No module named 'tensorflow_text'",
|
||||
"output_type": "error",
|
||||
"traceback": [
|
||||
"\u001b[1;31m---------------------------------------------------------------------------\u001b[0m",
|
||||
"\u001b[1;31mModuleNotFoundError\u001b[0m Traceback (most recent call last)",
|
||||
"\u001b[1;32m~\\AppData\\Local\\Temp/ipykernel_41180/4216669875.py\u001b[0m in \u001b[0;36m<module>\u001b[1;34m\u001b[0m\n\u001b[1;32m----> 1\u001b[1;33m \u001b[1;32mimport\u001b[0m \u001b[0mtensorflow_text\u001b[0m\u001b[1;33m\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0m\u001b[0;32m 2\u001b[0m \u001b[1;32mimport\u001b[0m \u001b[0mtensorflow_hub\u001b[0m \u001b[1;32mas\u001b[0m \u001b[0mhub\u001b[0m\u001b[1;33m\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0;32m 3\u001b[0m \u001b[0mvectorizer\u001b[0m \u001b[1;33m=\u001b[0m \u001b[0mhub\u001b[0m\u001b[1;33m.\u001b[0m\u001b[0mKerasLayer\u001b[0m\u001b[1;33m(\u001b[0m\u001b[1;34m'https://tfhub.dev/tensorflow/bert_en_uncased_preprocess/3'\u001b[0m\u001b[1;33m)\u001b[0m\u001b[1;33m\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n",
|
||||
"\u001b[1;31mModuleNotFoundError\u001b[0m: No module named 'tensorflow_text'"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"import tensorflow_text \n",
|
||||
"import tensorflow_hub as hub\n",
|
||||
"vectorizer = hub.KerasLayer('https://tfhub.dev/tensorflow/bert_en_uncased_preprocess/3')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"{'input_type_ids': <tf.Tensor: shape=(1, 128), dtype=int32, numpy=\n",
|
||||
" array([[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]],\n",
|
||||
" dtype=int32)>,\n",
|
||||
" 'input_word_ids': <tf.Tensor: shape=(1, 128), dtype=int32, numpy=\n",
|
||||
" array([[ 101, 1045, 2293, 19081, 102, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0]], dtype=int32)>,\n",
|
||||
" 'input_mask': <tf.Tensor: shape=(1, 128), dtype=int32, numpy=\n",
|
||||
" array([[1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]],\n",
|
||||
" dtype=int32)>}"
|
||||
]
|
||||
},
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"vectorizer(['I love transformers'])"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"使用與原始網絡訓練時相同的向量化工具是非常重要的。此外，BERT 向量化工具會返回三個組件：\n",
|
||||
"* `input_word_ids`，這是一個輸入句子的標記編號序列\n",
|
||||
"* `input_mask`，顯示序列中哪些部分包含實際輸入，哪些部分是填充。這與 `Masking` 層生成的遮罩類似\n",
|
||||
"* `input_type_ids` 用於語言建模任務，允許在一個序列中指定兩個輸入句子。\n",
|
||||
"\n",
|
||||
"然後，我們可以實例化 BERT 特徵提取器：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 8,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"bert = hub.KerasLayer('https://tfhub.dev/tensorflow/small_bert/bert_en_uncased_L-4_H-128_A-2/1')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 9,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"pooled_output -> (1, 128)\n",
|
||||
"encoder_outputs -> 4\n",
|
||||
"sequence_output -> (1, 128, 128)\n",
|
||||
"default -> (1, 128)\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"z = bert(vectorizer(['I love transformers']))\n",
|
||||
"for i,x in z.items():\n",
|
||||
" print(f\"{i} -> { len(x) if isinstance(x, list) else x.shape }\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"所以，BERT 層會返回一些有用的結果：\n",
|
||||
"* `pooled_output` 是通過平均序列中所有 token 的結果。你可以將其視為整個網絡的智能語義嵌入。它等同於我們之前模型中的 `GlobalAveragePooling1D` 層的輸出。\n",
|
||||
"* `sequence_output` 是最後一層 transformer 的輸出（對應於我們上面模型中的 `TransformerBlock` 的輸出）。\n",
|
||||
"* `encoder_outputs` 是所有 transformer 層的輸出。由於我們載入了 4 層的 BERT 模型（從名稱中包含的 `4_H` 你可能已經猜到），它有 4 個張量。最後一個張量與 `sequence_output` 相同。\n",
|
||||
"\n",
|
||||
"現在我們將定義端到端的分類模型。我們會使用*函數式模型定義*，即定義模型輸入，然後提供一系列表達式來計算其輸出。我們還會將 BERT 模型的權重設置為不可訓練，只訓練最終的分類器：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 10,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Model: \"model\"\n",
|
||||
"__________________________________________________________________________________________________\n",
|
||||
"Layer (type) Output Shape Param # Connected to \n",
|
||||
"==================================================================================================\n",
|
||||
"input_1 (InputLayer) [(None,)] 0 \n",
|
||||
"__________________________________________________________________________________________________\n",
|
||||
"keras_layer (KerasLayer) {'input_type_ids': ( 0 input_1[0][0] \n",
|
||||
"__________________________________________________________________________________________________\n",
|
||||
"keras_layer_1 (KerasLayer) {'pooled_output': (N 4782465 keras_layer[0][0] \n",
|
||||
" keras_layer[0][1] \n",
|
||||
" keras_layer[0][2] \n",
|
||||
"__________________________________________________________________________________________________\n",
|
||||
"dropout_4 (Dropout) (None, 128) 0 keras_layer_1[0][5] \n",
|
||||
"__________________________________________________________________________________________________\n",
|
||||
"dense_4 (Dense) (None, 4) 516 dropout_4[0][0] \n",
|
||||
"==================================================================================================\n",
|
||||
"Total params: 4,782,981\n",
|
||||
"Trainable params: 516\n",
|
||||
"Non-trainable params: 4,782,465\n",
|
||||
"__________________________________________________________________________________________________\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"inp = keras.Input(shape=(),dtype=tf.string)\n",
|
||||
"x = vectorizer(inp)\n",
|
||||
"x = bert(x)\n",
|
||||
"x = keras.layers.Dropout(0.1)(x['pooled_output'])\n",
|
||||
"out = keras.layers.Dense(4,activation='softmax')(x)\n",
|
||||
"model = keras.models.Model(inp,out)\n",
|
||||
"bert.trainable = False\n",
|
||||
"model.summary()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 11,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"938/938 [==============================] - 528s 559ms/step - loss: 0.8056 - acc: 0.6983 - val_loss: 0.5953 - val_acc: 0.7888\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<tensorflow.python.keras.callbacks.History at 0x7f9bb1e36d00>"
|
||||
]
|
||||
},
|
||||
"execution_count": 11,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model.compile(loss='sparse_categorical_crossentropy',metrics=['acc'], optimizer='adam')\n",
|
||||
"model.fit(ds_train.map(tupelize).batch(128),validation_data=ds_test.map(tupelize).batch(128))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"儘管可訓練的參數很少，但過程仍然相當緩慢，因為 BERT 特徵提取器的計算量很大。看起來我們無法達到合理的準確度，可能是因為訓練不足，或者模型參數不足。\n",
|
||||
"\n",
|
||||
"讓我們嘗試解凍 BERT 的權重並一起訓練。這需要非常小的學習率，還需要更謹慎的訓練策略，包括使用 **warmup** 和 **AdamW** 優化器。我們將使用 `tf-models-official` 套件來創建優化器：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 12,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Model: \"model\"\n",
|
||||
"__________________________________________________________________________________________________\n",
|
||||
"Layer (type) Output Shape Param # Connected to \n",
|
||||
"==================================================================================================\n",
|
||||
"input_1 (InputLayer) [(None,)] 0 \n",
|
||||
"__________________________________________________________________________________________________\n",
|
||||
"keras_layer (KerasLayer) {'input_type_ids': ( 0 input_1[0][0] \n",
|
||||
"__________________________________________________________________________________________________\n",
|
||||
"keras_layer_1 (KerasLayer) {'pooled_output': (N 4782465 keras_layer[0][0] \n",
|
||||
" keras_layer[0][1] \n",
|
||||
" keras_layer[0][2] \n",
|
||||
"__________________________________________________________________________________________________\n",
|
||||
"dropout_4 (Dropout) (None, 128) 0 keras_layer_1[0][5] \n",
|
||||
"__________________________________________________________________________________________________\n",
|
||||
"dense_4 (Dense) (None, 4) 516 dropout_4[0][0] \n",
|
||||
"==================================================================================================\n",
|
||||
"Total params: 4,782,981\n",
|
||||
"Trainable params: 4,782,980\n",
|
||||
"Non-trainable params: 1\n",
|
||||
"__________________________________________________________________________________________________\n",
|
||||
"938/938 [==============================] - 629s 664ms/step - loss: 0.6344 - acc: 0.7658 - val_loss: 0.4876 - val_acc: 0.8247\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<tensorflow.python.keras.callbacks.History at 0x7f9bb0bd0070>"
|
||||
]
|
||||
},
|
||||
"execution_count": 12,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"from official.nlp import optimization \n",
|
||||
"bert.trainable=True\n",
|
||||
"model.summary()\n",
|
||||
"epochs = 3\n",
|
||||
"opt = optimization.create_optimizer(\n",
|
||||
" init_lr=3e-5,\n",
|
||||
" num_train_steps=epochs*len(ds_train),\n",
|
||||
" num_warmup_steps=0.1*epochs*len(ds_train),\n",
|
||||
" optimizer_type='adamw')\n",
|
||||
"\n",
|
||||
"model.compile(loss='sparse_categorical_crossentropy',metrics=['acc'], optimizer=opt)\n",
|
||||
"model.fit(ds_train.map(tupelize).batch(128),validation_data=ds_test.map(tupelize).batch(128))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"如你所見，訓練過程相當緩慢——但你可能想嘗試進行幾個訓練週期（5-10次），看看是否能夠與我們之前使用的方法相比，獲得最佳結果。\n",
|
||||
"\n",
|
||||
"## Huggingface Transformers 庫\n",
|
||||
"\n",
|
||||
"另一種非常常見（且稍微簡單）的使用 Transformer 模型的方法是 [HuggingFace 套件](https://github.com/huggingface/)，它為不同的 NLP 任務提供了簡單的構建模塊。此套件同時支援 Tensorflow 和 PyTorch，後者是另一個非常受歡迎的神經網絡框架。\n",
|
||||
"\n",
|
||||
"> **注意**：如果你對了解 Transformers 庫的運作方式不感興趣——你可以跳到筆記本的最後部分，因為你不會看到任何與我們之前所做的有實質性不同的內容。我們將重複使用不同的庫和更大的模型來訓練 BERT 模型的相同步驟。因此，這個過程涉及一些相當長的訓練時間，所以你可能只想瀏覽一下程式碼。\n",
|
||||
"\n",
|
||||
"讓我們看看如何使用 [Huggingface Transformers](http://huggingface.co) 解決我們的問題。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"首先，我們需要選擇要使用的模型。除了內建的模型之外，Huggingface 還有一個[線上模型庫](https://huggingface.co/models)，社群提供了許多預訓練模型。只需提供模型名稱，就可以載入並使用這些模型。所有模型所需的二進制檔案會自動下載。\n",
|
||||
"\n",
|
||||
"有時候你可能需要載入自己的模型，這時可以指定包含所有相關檔案的目錄，包括 tokenizer 的參數、`config.json` 文件（包含模型參數）、二進制權重等。\n",
|
||||
"\n",
|
||||
"透過模型名稱，我們可以初始化模型和 tokenizer。讓我們先從 tokenizer 開始：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 2,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import transformers\n",
|
||||
"\n",
|
||||
"# To load the model from Internet repository using model name. \n",
|
||||
"# Use this if you are running from your own copy of the notebooks\n",
|
||||
"bert_model = 'bert-base-uncased' \n",
|
||||
"\n",
|
||||
"# To load the model from the directory on disk. Use this for Microsoft Learn module, because we have\n",
|
||||
"# prepared all required files for you.\n",
|
||||
"#bert_model = './bert'\n",
|
||||
"\n",
|
||||
"tokenizer = transformers.BertTokenizer.from_pretrained(bert_model)\n",
|
||||
"\n",
|
||||
"MAX_SEQ_LEN = 128\n",
|
||||
"PAD_INDEX = tokenizer.convert_tokens_to_ids(tokenizer.pad_token)\n",
|
||||
"UNK_INDEX = tokenizer.convert_tokens_to_ids(tokenizer.unk_token)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"`tokenizer` 對象包含可直接用於編碼文本的 `encode` 函數：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"[101, 23435, 12314, 2003, 1037, 2307, 7705, 2005, 17953, 2361, 102]"
|
||||
]
|
||||
},
|
||||
"execution_count": 3,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"tokenizer.encode('Tensorflow is a great framework for NLP')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們還可以使用分詞器以適合傳遞給模型的方式對序列進行編碼，即包括 `token_ids`、`input_mask` 等字段。我們還可以通過提供 `return_tensors='tf'` 參數來指定我們想要 Tensorflow 張量：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"{'input_ids': <tf.Tensor: shape=(1, 5), dtype=int32, numpy=array([[ 101, 7592, 1010, 2045, 102]], dtype=int32)>, 'token_type_ids': <tf.Tensor: shape=(1, 5), dtype=int32, numpy=array([[0, 0, 0, 0, 0]], dtype=int32)>, 'attention_mask': <tf.Tensor: shape=(1, 5), dtype=int32, numpy=array([[1, 1, 1, 1, 1]], dtype=int32)>}"
|
||||
]
|
||||
},
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"tokenizer(['Hello, there'],return_tensors='tf')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"在我們的案例中，我們將使用預訓練的 BERT 模型，名為 `bert-base-uncased`。*Uncased* 表示該模型對大小寫不敏感。\n",
|
||||
"\n",
|
||||
"在訓練模型時，我們需要提供已分詞的序列作為輸入，因此我們將設計數據處理管道。由於 `tokenizer.encode` 是一個 Python 函數，我們將採用與上一單元相同的方法，使用 `py_function` 來調用它：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 31,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"def process(x):\n",
|
||||
" return tokenizer.encode(x.numpy().decode('utf-8'),return_tensors='tf',padding='max_length',max_length=MAX_SEQ_LEN,truncation=True)[0]\n",
|
||||
"\n",
|
||||
"def process_fn(x):\n",
|
||||
" s = x['title']+' '+x['description']\n",
|
||||
" e = tf.py_function(process,inp=[s],Tout=(tf.int32))\n",
|
||||
" e.set_shape(MAX_SEQ_LEN)\n",
|
||||
" return e,x['label']"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在我們可以使用 `BertForSequenceClassfication` 套件加載實際模型。這確保了我們的模型已經具備分類所需的架構，包括最終的分類器。你會看到一條警告訊息，指出最終分類器的權重尚未初始化，並且模型需要進行預訓練——這完全沒問題，因為這正是我們即將進行的操作！\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 32,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"model = transformers.TFBertForSequenceClassification.from_pretrained(bert_model,num_labels=4,output_attentions=False)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 33,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Model: \"tf_bert_for_sequence_classification_1\"\n",
|
||||
"_________________________________________________________________\n",
|
||||
"Layer (type) Output Shape Param # \n",
|
||||
"=================================================================\n",
|
||||
"bert (TFBertMainLayer) multiple 109482240 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"dropout_75 (Dropout) multiple 0 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"classifier (Dense) multiple 3076 \n",
|
||||
"=================================================================\n",
|
||||
"Total params: 109,485,316\n",
|
||||
"Trainable params: 109,485,316\n",
|
||||
"Non-trainable params: 0\n",
|
||||
"_________________________________________________________________\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model.summary()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"從 `summary()` 可以看到，該模型包含了接近 1.1 億個參數！假設我們想在相對較小的數據集上進行簡單的分類任務，我們可能不希望訓練 BERT 的基礎層：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 34,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Model: \"tf_bert_for_sequence_classification_1\"\n",
|
||||
"_________________________________________________________________\n",
|
||||
"Layer (type) Output Shape Param # \n",
|
||||
"=================================================================\n",
|
||||
"bert (TFBertMainLayer) multiple 109482240 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"dropout_75 (Dropout) multiple 0 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"classifier (Dense) multiple 3076 \n",
|
||||
"=================================================================\n",
|
||||
"Total params: 109,485,316\n",
|
||||
"Trainable params: 3,076\n",
|
||||
"Non-trainable params: 109,482,240\n",
|
||||
"_________________________________________________________________\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model.layers[0].trainable = False\n",
|
||||
"model.summary()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在我們準備開始訓練！\n",
|
||||
"\n",
|
||||
"> **注意**：訓練完整規模的 BERT 模型可能會非常耗時！因此，我們只會訓練前 32 個批次。這只是為了展示模型訓練的設置方式。如果你有興趣嘗試完整規模的訓練，只需移除 `steps_per_epoch` 和 `validation_steps` 參數，然後準備耐心等待！\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 30,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"32/32 [==============================] - 142s 4s/step - loss: 1.3896 - acc: 0.2500 - val_loss: 1.3863 - val_acc: 0.2480\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<tensorflow.python.keras.callbacks.History at 0x7f1d40a4b6a0>"
|
||||
]
|
||||
},
|
||||
"execution_count": 30,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model.compile('adam','sparse_categorical_crossentropy',['acc'])\n",
|
||||
"tf.get_logger().setLevel('ERROR')\n",
|
||||
"model.fit(ds_train.map(process_fn).batch(32),validation_data=ds_test.map(process_fn).batch(32),steps_per_epoch=32,validation_steps=2)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"如果你增加迭代次數並耐心等待，並且訓練多個epoch，你可以預期BERT分類會給我們最好的準確率！這是因為BERT已經相當理解語言的結構，我們只需要微調最終的分類器。然而，由於BERT是一個大型模型，整個訓練過程需要很長時間，並且需要強大的計算能力！（GPU，最好是多於一個）。\n",
|
||||
"\n",
|
||||
"> **Note:** 在我們的例子中，我們使用的是其中一個最小的預訓練BERT模型。還有更大的模型可能會帶來更好的結果。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 重點\n",
|
||||
"\n",
|
||||
"在本單元中，我們探討了基於 **transformers** 的最新模型架構。我們已將其應用於文本分類任務，但同樣地，BERT 模型也可以用於實體抽取、問題回答以及其他自然語言處理任務。\n",
|
||||
"\n",
|
||||
"Transformer 模型代表了自然語言處理領域的最新技術，並且在大多數情況下，應該是您在實現自定義 NLP 解決方案時首先嘗試的選擇。然而，如果您希望構建更高級的神經網絡模型，理解本模組中討論的循環神經網絡的基本原理是非常重要的。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": []
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件已使用人工智能翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。應以原文文件作為權威來源。對於關鍵資訊，建議尋求專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或誤釋不承擔責任。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"interpreter": {
|
||||
"hash": "0cb620c6d4b9f7a635928804c26cf22403d89d98d79684e4529119355ee6d5a5"
|
||||
},
|
||||
"kernelspec": {
|
||||
"display_name": "py38_tensorflow",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.8.12"
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "ab59c532409774988ab875f2260e8e53",
|
||||
"translation_date": "2025-08-31T10:43:28+00:00",
|
||||
"source_file": "lessons/5-NLP/18-Transformers/TransformersTF.ipynb",
|
||||
"language_code": "hk"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 4
|
||||
}
|
||||
|
|
@ -0,0 +1,492 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 命名實體識別 (NER)\n",
|
||||
"\n",
|
||||
"此筆記本來自 [AI for Beginners Curriculum](http://aka.ms/ai-beginners)。\n",
|
||||
"\n",
|
||||
"在這個例子中，我們將學習如何在 [Annotated Corpus for Named Entity Recognition](https://www.kaggle.com/datasets/abhinavwalia95/entity-annotated-corpus) 資料集上訓練 NER 模型。在開始之前，請下載 [ner_dataset.csv](https://www.kaggle.com/datasets/abhinavwalia95/entity-annotated-corpus?resource=download&select=ner_dataset.csv) 文件到當前目錄。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 62,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import pandas as pd\n",
|
||||
"from tensorflow import keras\n",
|
||||
"import numpy as np"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 準備數據集\n",
|
||||
"\n",
|
||||
"我們將從將數據集讀入一個 dataframe 開始。如果你想了解更多有關使用 Pandas 的資訊，可以參考我們 [初學者數據科學](http://aka.ms/datascience-beginners) 中的 [數據處理課程](https://github.com/microsoft/Data-Science-For-Beginners/tree/main/2-Working-With-Data/07-python)。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/html": [
|
||||
"<div>\n",
|
||||
"<style scoped>\n",
|
||||
" .dataframe tbody tr th:only-of-type {\n",
|
||||
" vertical-align: middle;\n",
|
||||
" }\n",
|
||||
"\n",
|
||||
" .dataframe tbody tr th {\n",
|
||||
" vertical-align: top;\n",
|
||||
" }\n",
|
||||
"\n",
|
||||
" .dataframe thead th {\n",
|
||||
" text-align: right;\n",
|
||||
" }\n",
|
||||
"</style>\n",
|
||||
"<table border=\"1\" class=\"dataframe\">\n",
|
||||
" <thead>\n",
|
||||
" <tr style=\"text-align: right;\">\n",
|
||||
" <th></th>\n",
|
||||
" <th>Sentence #</th>\n",
|
||||
" <th>Word</th>\n",
|
||||
" <th>POS</th>\n",
|
||||
" <th>Tag</th>\n",
|
||||
" </tr>\n",
|
||||
" </thead>\n",
|
||||
" <tbody>\n",
|
||||
" <tr>\n",
|
||||
" <th>0</th>\n",
|
||||
" <td>Sentence: 1</td>\n",
|
||||
" <td>Thousands</td>\n",
|
||||
" <td>NNS</td>\n",
|
||||
" <td>O</td>\n",
|
||||
" </tr>\n",
|
||||
" <tr>\n",
|
||||
" <th>1</th>\n",
|
||||
" <td>NaN</td>\n",
|
||||
" <td>of</td>\n",
|
||||
" <td>IN</td>\n",
|
||||
" <td>O</td>\n",
|
||||
" </tr>\n",
|
||||
" <tr>\n",
|
||||
" <th>2</th>\n",
|
||||
" <td>NaN</td>\n",
|
||||
" <td>demonstrators</td>\n",
|
||||
" <td>NNS</td>\n",
|
||||
" <td>O</td>\n",
|
||||
" </tr>\n",
|
||||
" <tr>\n",
|
||||
" <th>3</th>\n",
|
||||
" <td>NaN</td>\n",
|
||||
" <td>have</td>\n",
|
||||
" <td>VBP</td>\n",
|
||||
" <td>O</td>\n",
|
||||
" </tr>\n",
|
||||
" <tr>\n",
|
||||
" <th>4</th>\n",
|
||||
" <td>NaN</td>\n",
|
||||
" <td>marched</td>\n",
|
||||
" <td>VBN</td>\n",
|
||||
" <td>O</td>\n",
|
||||
" </tr>\n",
|
||||
" </tbody>\n",
|
||||
"</table>\n",
|
||||
"</div>"
|
||||
],
|
||||
"text/plain": [
|
||||
" Sentence # Word POS Tag\n",
|
||||
"0 Sentence: 1 Thousands NNS O\n",
|
||||
"1 NaN of IN O\n",
|
||||
"2 NaN demonstrators NNS O\n",
|
||||
"3 NaN have VBP O\n",
|
||||
"4 NaN marched VBN O"
|
||||
]
|
||||
},
|
||||
"execution_count": 3,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"df = pd.read_csv('ner_dataset.csv',encoding='unicode-escape')\n",
|
||||
"df.head()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"讓我們獲取獨特的標籤並創建查詢字典，以便我們可以用來將標籤轉換為類別編號：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"array(['O', 'B-geo', 'B-gpe', 'B-per', 'I-geo', 'B-org', 'I-org', 'B-tim',\n",
|
||||
" 'B-art', 'I-art', 'I-per', 'I-gpe', 'I-tim', 'B-nat', 'B-eve',\n",
|
||||
" 'I-eve', 'I-nat'], dtype=object)"
|
||||
]
|
||||
},
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"tags = df.Tag.unique()\n",
|
||||
"tags"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 8,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"'O'"
|
||||
]
|
||||
},
|
||||
"execution_count": 8,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"id2tag = dict(enumerate(tags))\n",
|
||||
"tag2id = { v : k for k,v in id2tag.items() }\n",
|
||||
"\n",
|
||||
"id2tag[0]"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在我們需要對詞彙做同樣的處理。為了簡化，我們將在不考慮詞頻的情況下創建詞彙；在現實中，你可能會想使用 Keras 向量化工具，並限制詞彙的數量。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 14,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"vocab = set(df['Word'].apply(lambda x: x.lower()))\n",
|
||||
"id2word = { i+1 : v for i,v in enumerate(vocab) }\n",
|
||||
"id2word[0] = '<UNK>'\n",
|
||||
"vocab.add('<UNK>')\n",
|
||||
"word2id = { v : k for k,v in id2word.items() }"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們需要建立一個句子數據集用於訓練。讓我們遍歷原始數據集，並將所有單個句子分成 `X`（單詞列表）和 `Y`（標記列表）：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 41,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"X,Y = [],[]\n",
|
||||
"s,t = [],[]\n",
|
||||
"for i,row in df[['Sentence #','Word','Tag']].iterrows():\n",
|
||||
" if pd.isna(row['Sentence #']):\n",
|
||||
" s.append(row['Word'])\n",
|
||||
" t.append(row['Tag'])\n",
|
||||
" else:\n",
|
||||
" if len(s)>0:\n",
|
||||
" X.append(s)\n",
|
||||
" Y.append(t)\n",
|
||||
" s,t = [row['Word']],[row['Tag']]\n",
|
||||
"X.append(s)\n",
|
||||
"Y.append(t)\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": []
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 93,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"([10386,\n",
|
||||
" 23515,\n",
|
||||
" 4134,\n",
|
||||
" 29620,\n",
|
||||
" 7954,\n",
|
||||
" 13583,\n",
|
||||
" 21193,\n",
|
||||
" 12222,\n",
|
||||
" 27322,\n",
|
||||
" 18258,\n",
|
||||
" 5815,\n",
|
||||
" 15880,\n",
|
||||
" 5355,\n",
|
||||
" 25242,\n",
|
||||
" 31327,\n",
|
||||
" 18258,\n",
|
||||
" 27067,\n",
|
||||
" 23515,\n",
|
||||
" 26444,\n",
|
||||
" 14412,\n",
|
||||
" 358,\n",
|
||||
" 26551,\n",
|
||||
" 5011,\n",
|
||||
" 30558],\n",
|
||||
" [0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0])"
|
||||
]
|
||||
},
|
||||
"execution_count": 93,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"def vectorize(seq):\n",
|
||||
" return [word2id[x.lower()] for x in seq]\n",
|
||||
"\n",
|
||||
"def tagify(seq):\n",
|
||||
" return [tag2id[x] for x in seq]\n",
|
||||
"\n",
|
||||
"Xv = list(map(vectorize,X))\n",
|
||||
"Yv = list(map(tagify,Y))\n",
|
||||
"\n",
|
||||
"Xv[0], Yv[0]"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"為了簡化，我們將所有句子用 0 代幣填充到最大長度。在現實生活中，我們可能會想使用更聰明的策略，僅在一個小批次內填充序列。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 51,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"X_data = keras.preprocessing.sequence.pad_sequences(Xv,padding='post')\n",
|
||||
"Y_data = keras.preprocessing.sequence.pad_sequences(Yv,padding='post')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 定義標記分類網絡\n",
|
||||
"\n",
|
||||
"我們將使用兩層雙向 LSTM 網絡進行標記分類。為了將密集分類器應用於最後一層 LSTM 的每個輸出，我們會使用 `TimeDistributed` 結構，該結構會在每一步中將相同的密集層複製到 LSTM 的每個輸出：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 94,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Model: \"sequential_3\"\n",
|
||||
"_________________________________________________________________\n",
|
||||
" Layer (type) Output Shape Param # \n",
|
||||
"=================================================================\n",
|
||||
" embedding_4 (Embedding) (None, 104, 300) 9545400 \n",
|
||||
" \n",
|
||||
" bidirectional_6 (Bidirectio (None, 104, 200) 320800 \n",
|
||||
" nal) \n",
|
||||
" \n",
|
||||
" bidirectional_7 (Bidirectio (None, 104, 200) 240800 \n",
|
||||
" nal) \n",
|
||||
" \n",
|
||||
" time_distributed_3 (TimeDis (None, 104, 17) 3417 \n",
|
||||
" tributed) \n",
|
||||
" \n",
|
||||
"=================================================================\n",
|
||||
"Total params: 10,110,417\n",
|
||||
"Trainable params: 10,110,417\n",
|
||||
"Non-trainable params: 0\n",
|
||||
"_________________________________________________________________\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"maxlen = X_data.shape[1]\n",
|
||||
"vocab_size = len(vocab)\n",
|
||||
"num_tags = len(tags)\n",
|
||||
"model = keras.models.Sequential([\n",
|
||||
" keras.layers.Embedding(vocab_size, 300, input_length=maxlen),\n",
|
||||
" keras.layers.Bidirectional(keras.layers.LSTM(units=100, activation='tanh', return_sequences=True)),\n",
|
||||
" keras.layers.Bidirectional(keras.layers.LSTM(units=100, activation='tanh', return_sequences=True)),\n",
|
||||
" keras.layers.TimeDistributed(keras.layers.Dense(num_tags, activation='softmax'))\n",
|
||||
"])\n",
|
||||
"model.compile(loss='sparse_categorical_crossentropy',optimizer='adam',metrics=['acc'])\n",
|
||||
"model.summary()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"注意這裡，我們為數據集明確指定了 `maxlen`——如果我們希望網絡能夠處理可變長度的序列，那麼在定義網絡時需要更聰明一些。\n",
|
||||
"\n",
|
||||
"現在讓我們訓練模型。為了加快速度，我們只訓練一個 epoch，但你可以嘗試訓練更長的時間。此外，你可能希望將部分數據集分離作為訓練數據集，以觀察驗證準確率。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 57,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"1499/1499 [==============================] - 740s 488ms/step - loss: 0.0667 - acc: 0.9841\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<keras.callbacks.History at 0x16f0bb2a310>"
|
||||
]
|
||||
},
|
||||
"execution_count": 57,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model.fit(X_data,Y_data)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 測試結果\n",
|
||||
"\n",
|
||||
"現在讓我們看看實體識別模型在一個示例句子上的表現：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 91,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"sent = 'John Smith went to Paris to attend a conference in cancer development institute'\n",
|
||||
"words = sent.lower().split()\n",
|
||||
"v = keras.preprocessing.sequence.pad_sequences([[word2id[x] for x in words]],padding='post',maxlen=maxlen)\n",
|
||||
"res = model(v)[0]"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 92,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"john -> B-per\n",
|
||||
"smith -> I-per\n",
|
||||
"went -> O\n",
|
||||
"to -> O\n",
|
||||
"paris -> B-geo\n",
|
||||
"to -> O\n",
|
||||
"attend -> O\n",
|
||||
"a -> O\n",
|
||||
"conference -> O\n",
|
||||
"in -> O\n",
|
||||
"cancer -> B-org\n",
|
||||
"development -> I-org\n",
|
||||
"institute -> I-org\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"r = np.argmax(res.numpy(),axis=1)\n",
|
||||
"for i,w in zip(r,words):\n",
|
||||
" print(f\"{w} -> {id2tag[i]}\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 重點\n",
|
||||
"\n",
|
||||
"即使是簡單的 LSTM 模型，在命名實體識別（NER）方面也能展示出不錯的效果。然而，若要獲得更佳的結果，你可能需要使用大型的預訓練語言模型，例如 BERT。使用 Huggingface Transformers 庫來訓練 BERT 用於 NER 的方法已在[這裡](https://huggingface.co/course/chapter7/2?fw=pt)進行了說明。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件已使用人工智能翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。應以原文文件作為權威來源。對於關鍵資訊，建議尋求專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或誤釋不承擔責任。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"interpreter": {
|
||||
"hash": "16af2a8bbb083ea23e5e41c7f5787656b2ce26968575d8763f2c4b17f9cd711f"
|
||||
},
|
||||
"kernelspec": {
|
||||
"display_name": "Python 3.8.12 ('py38')",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.8.12"
|
||||
},
|
||||
"orig_nbformat": 4,
|
||||
"coopTranslator": {
|
||||
"original_hash": "254d25052dcca4ef84f59a05f2935bdc",
|
||||
"translation_date": "2025-08-31T10:45:44+00:00",
|
||||
"source_file": "lessons/5-NLP/19-NER/NER-TF.ipynb",
|
||||
"language_code": "hk"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 2
|
||||
}
|
||||
|
|
@ -0,0 +1,325 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"attachments": {},
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 嘗試使用 OpenAI GPT\n",
|
||||
"\n",
|
||||
"此筆記本是 [AI 初學者課程](http://aka.ms/ai-beginners) 的一部分。\n",
|
||||
"\n",
|
||||
"在這個筆記本中，我們將探索如何使用 Hugging Face 的 `transformers` 庫來操作 OpenAI-GPT 模型。\n",
|
||||
"\n",
|
||||
"事不宜遲，讓我們建立文字生成管道並開始生成吧！\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 1,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stderr",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"c:\\Users\\bethanycheum\\Desktop\\AI-For-Beginners\\.venv\\lib\\site-packages\\tqdm\\auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html\n",
|
||||
" from .autonotebook import tqdm as notebook_tqdm\n",
|
||||
"Downloading model.safetensors: 100%|██████████| 479M/479M [04:28<00:00, 1.78MB/s] \n",
|
||||
"c:\\Users\\bethanycheum\\Desktop\\AI-For-Beginners\\.venv\\lib\\site-packages\\huggingface_hub\\file_download.py:133: UserWarning: `huggingface_hub` cache-system uses symlinks by default to efficiently store duplicated files but your machine does not support them in C:\\Users\\bethanycheum\\.cache\\huggingface\\hub. Caching files will still work but in a degraded version that might require more space on your disk. This warning can be disabled by setting the `HF_HUB_DISABLE_SYMLINKS_WARNING` environment variable. For more details, see https://huggingface.co/docs/huggingface_hub/how-to-cache#limitations.\n",
|
||||
"To support symlinks on Windows, you either need to activate Developer Mode or to run Python as an administrator. In order to see activate developer mode, see this article: https://docs.microsoft.com/en-us/windows/apps/get-started/enable-your-device-for-development\n",
|
||||
" warnings.warn(message)\n",
|
||||
"Some weights of OpenAIGPTLMHeadModel were not initialized from the model checkpoint at openai-gpt and are newly initialized: ['position_ids']\n",
|
||||
"You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\n",
|
||||
"Downloading (…)neration_config.json: 100%|██████████| 74.0/74.0 [00:00<00:00, 48.8kB/s]\n",
|
||||
"Downloading (…)olve/main/vocab.json: 100%|██████████| 816k/816k [00:00<00:00, 1.76MB/s]\n",
|
||||
"Downloading (…)olve/main/merges.txt: 100%|██████████| 458k/458k [00:00<00:00, 1.11MB/s]\n",
|
||||
"Downloading (…)/main/tokenizer.json: 100%|██████████| 1.27M/1.27M [00:00<00:00, 2.12MB/s]\n",
|
||||
"Xformers is not installed correctly. If you want to use memory_efficient_attention to accelerate training use the following command to install Xformers\n",
|
||||
"pip install xformers.\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"[{'generated_text': \"Hello! I am a neural network, and I want to say that i apologize for not coming to you yourself, for not helping you, and that i was too busy getting dressed and studying for a midterm. you know, the kind where the teachers are like that and they come in pairs with their boyfriends, but not with theirs. it's true, that i have had a girlfriend, and i'm only going on wednesdays and thursdays because i was too busy with college, but maybe\"},\n",
|
||||
" {'generated_text': 'Hello! I am a neural network, and I want to say that we have been blessed with a wonderful gift ; no one of us has died at all. and our spirits are strong, very strong. in one very lucky moment of luck for you, all has been given direction and destiny, and for us there are no more mysteries. the earth has been chosen for you, and that earth is now ours, and you must be forever in our hearts. \" \\n the words, as one,'},\n",
|
||||
" {'generated_text': 'Hello! I am a neural network, and I want to say that if you would just turn and face the general, you would have a nice day. \" \\n \" sure thing, \" said one of the soldiers, and started to run. the rest of the soldiers followed, shouting. the general turned to general zulu, raising his arm. the general said something in his native language, and the general immediately started to run. zulu started to move toward the wall, with the'},\n",
|
||||
" {'generated_text': 'Hello! I am a neural network, and I want to say that i am not a doctor but an anthropologist to you, a specialist, a specialist in the field of astrobiological biology, and that i am very much involved in this investigation. i am not sure, i am not certain, but i can confirm your conclusions and therefore i will go to the top. i have a colleague who has just returned from this expedition and his findings confirm that you are a specialist. that is, he'},\n",
|
||||
" {'generated_text': \"Hello! I am a neural network, and I want to say that everyone here is in agreement that no matter how many times i say to myself,'he was never a man of action on the battlefield,'or'he 'll never take a chance at killing any civilians,'or'he 'll never let his men go undefended against enemy forces of this caliber,'or'that's just what i need in a day like today. \\n you see, there are only three groups that\"}]"
|
||||
]
|
||||
},
|
||||
"execution_count": 1,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"from transformers import pipeline\n",
|
||||
"\n",
|
||||
"model_name = 'openai-gpt' \n",
|
||||
"\n",
|
||||
"generator = pipeline('text-generation', model=model_name)\n",
|
||||
"\n",
|
||||
"generator(\"Hello! I am a neural network, and I want to say that\", max_length=100, num_return_sequences=5)\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"attachments": {},
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 提示工程\n",
|
||||
"\n",
|
||||
"在某些問題中，你可以通過設計正確的提示直接使用 openai-gpt 生成。以下是一些示例：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 2,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"[{'generated_text': 'Synonyms of a word cat: the same cat i used to stare at, and you in'},\n",
|
||||
" {'generated_text': 'Synonyms of a word cat: cat of the woods, cat of the hills, cat of'},\n",
|
||||
" {'generated_text': 'Synonyms of a word cat: you! \\n \" it\\'s a girl. \" i said'},\n",
|
||||
" {'generated_text': \"Synonyms of a word cat: big cat. but how come, we didn't hear it\"},\n",
|
||||
" {'generated_text': 'Synonyms of a word cat: \" mea - o - c \" which makes them sound'}]"
|
||||
]
|
||||
},
|
||||
"execution_count": 2,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"generator(\"Synonyms of a word cat:\", max_length=20, num_return_sequences=5)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"[{'generated_text': 'I love when you say this -> Positive\\nI have myself -> Negative\\nThis is awful for you to say this -> positive this is so horrible - > positive that your brother is gay - >'},\n",
|
||||
" {'generated_text': 'I love when you say this -> Positive\\nI have myself -> Negative\\nThis is awful for you to say this -> negative i will bring this on you -, < positive am i, i'},\n",
|
||||
" {'generated_text': 'I love when you say this -> Positive\\nI have myself -> Negative\\nThis is awful for you to say this -> negative i have self - esteem i must take it - : \\n - -'},\n",
|
||||
" {'generated_text': 'I love when you say this -> Positive\\nI have myself -> Negative\\nThis is awful for you to say this -> negative this is - : \\n if it were true that the devil would have'},\n",
|
||||
" {'generated_text': \"I love when you say this -> Positive\\nI have myself -> Negative\\nThis is awful for you to say this -> positive i have you - > positive it's a bad thing, > positive\"}]"
|
||||
]
|
||||
},
|
||||
"execution_count": 3,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"generator(\"I love when you say this -> Positive\\nI have myself -> Negative\\nThis is awful for you to say this ->\", max_length=40, num_return_sequences=5)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"[{'generated_text': 'Translate English to French: cat => chat, dog => chien, student => new and unusual. there were no more words to be'},\n",
|
||||
" {'generated_text': 'Translate English to French: cat => chat, dog => chien, student => student \\n his eyes were huge in his lean face as'},\n",
|
||||
" {'generated_text': \"Translate English to French: cat => chat, dog => chien, student => the teacher's words, their words, their words.\"}]"
|
||||
]
|
||||
},
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"generator(\"Translate English to French: cat => chat, dog => chien, student => \", top_k=50, max_length=30, num_return_sequences=3)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"[{'generated_text': 'People who liked the movie The Matrix also liked it, and there was the movie of the first man after us. \\n i wanted to laugh at how stupid these stupid actors were. no, they were'},\n",
|
||||
" {'generated_text': \"People who liked the movie The Matrix also liked the movie, and the film was the result. and that's when the man in the story was brought into reality, after a few decades. \\n a\"},\n",
|
||||
" {'generated_text': 'People who liked the movie The Matrix also liked the movie the matrix, because there was a very old movie movie called the matrix, where there was a great super hero, and the super hero came out'},\n",
|
||||
" {'generated_text': \"People who liked the movie The Matrix also liked the movie that didn't have a chance to pay cash, if they could afford it. most often they got a good deal and a lot of money,\"},\n",
|
||||
" {'generated_text': \"People who liked the movie The Matrix also liked the movie, and i didn't seem to have the same problem. \\n i 'd met the other half of my family. i spent most of my time\"}]"
|
||||
]
|
||||
},
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"generator(\"People who liked the movie The Matrix also liked \", max_length=40, num_return_sequences=5)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 文本取樣策略\n",
|
||||
"\n",
|
||||
"到目前為止，我們一直使用簡單的 **貪婪** 取樣策略，根據最高概率選擇下一個詞語。以下是其運作方式：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 6,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"[{'generated_text': 'It was early evening when I can back from work. I usually work late, but this time it was an exception. When I entered a room, I saw my friend, a young man, sprawled across the bed in his bed. \\n \" hi, i\\'m mike eptirard. \" \\n there was silence on the other side of the door. i listened for any trace of life but there was nothing. my heart began to pound, i was starting to sweat, i took out my wallet'},\n",
|
||||
" {'generated_text': 'It was early evening when I can back from work. I usually work late, but this time it was an exception. When I entered a room, I saw my mother on the bed, hugging her legs to her chest and sobbing. i saw my dad and mother from the corner of my eye. \\n elfin face was covered in tears as i entered the room. my dad and mother also wept ; just as they did every other time i came to work. but this time, they had different faces'},\n",
|
||||
" {'generated_text': 'It was early evening when I can back from work. I usually work late, but this time it was an exception. When I entered a room, I saw the room had changed because it was dark. it still smelled like a hospital. a new light shined through from a vent in the ceiling. i found myself in a bathroom and a small room with a sink and a wall of glass. the bathroom billion years ago. not so different from all of the rest of the apartment. \\n now...'},\n",
|
||||
" {'generated_text': 'It was early evening when I can back from work. I usually work late, but this time it was an exception. When I entered a room, I saw a large woman with dark hair and pale skin. she was asleep, but i noticed a faint movement of her face. i could sense she was awake. i got up and walked over to her. \\n \" hello miss. i am inspector michael o\\'dell ; we are investigating the case against you. i wanted to ask if you were the'},\n",
|
||||
" {'generated_text': 'It was early evening when I can back from work. I usually work late, but this time it was an exception. When I entered a room, I saw i had an empty table and three empty chairs. that was all i needed. i had left a note on a table in the center of the room and had a pen in hand. \" \\n \" i think what you were doing was something he was doing to her. \" \\n \" yeah, \" i nodded with a grin. \" i'}]"
|
||||
]
|
||||
},
|
||||
"execution_count": 6,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"prompt = \"It was early evening when I can back from work. I usually work late, but this time it was an exception. When I entered a room, I saw\"\n",
|
||||
"generator(prompt,max_length=100,num_return_sequences=5)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"**束搜索**允許生成器探索多個文本生成的方向（*束*），並選擇整體分數較高的方向。您可以通過提供 `num_beams` 參數來進行束搜索。您還可以指定 `no_repeat_ngram_size`，以懲罰模型重複給定大小的 n-gram：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"[{'generated_text': 'It was early evening when I can back from work. I usually work late, but this time it was an exception. When I entered a room, I saw a man sitting in a chair with his head in his hands. he didn\\'t look up as i approached. \\n \" excuse me, sir, \" i said. \" can i help you? \" \\n the man looked up at me. his eyes were red - rimmed and his face was pale, as if he hadn\\'t slept in days'},\n",
|
||||
" {'generated_text': 'It was early evening when I can back from work. I usually work late, but this time it was an exception. When I entered a room, I saw a man sitting at a desk in the middle of the room. he had his back to me, so i couldn\\'t see what he was doing. \" \\n \" what did he look like? \" i asked as i sat down on the bed next to her. \\n she took a deep breath and looked at me with tears in her eyes'},\n",
|
||||
" {'generated_text': 'It was early evening when I can back from work. I usually work late, but this time it was an exception. When I entered a room, I saw a woman sitting on the bed, reading a book. she looked up at me and smiled. \\n \" hi, \" she said. \" can i help you? \" \\n i sat down next to her and looked around the room. the walls were white, and there was a large window in the middle of the wall that looked out on'},\n",
|
||||
" {'generated_text': 'It was early evening when I can back from work. I usually work late, but this time it was an exception. When I entered a room, I saw a man sitting at a table in the middle of the room. he looked up as i walked in, and when he saw me, he got up and walked over to me. \\n \" can i help you? \" he asked as he put his hand on the small of my back and led me to a chair at the other end of'},\n",
|
||||
" {'generated_text': 'It was early evening when I can back from work. I usually work late, but this time it was an exception. When I entered a room, I saw a woman sitting on the edge of her bed, reading a book. she looked up at me and smiled. \\n \" hello, \" she said. \" can i help you? \" \\n i didn\\'t know what to say, so i just sat down in the chair next to the bed and looked at her. her hair was dark brown'}]"
|
||||
]
|
||||
},
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"prompt = \"It was early evening when I can back from work. I usually work late, but this time it was an exception. When I entered a room, I saw\"\n",
|
||||
"generator(prompt,max_length=100,num_return_sequences=5,num_beams=10,no_repeat_ngram_size=2)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"**抽樣** 根據模型返回的概率分佈，以非確定性的方式選擇下一個詞語。您可以使用參數 `do_sample=True` 開啟抽樣功能。您還可以指定 `temperature`，以調整模型的確定性程度。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 8,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"[{'generated_text': 'It was early evening when I can back from work. I usually work late, but this time it was an exception. When I entered a room, I saw her. she was on the bed, but she looked very different. \\n \" honey, what\\'s the matter? \" i asked. \\n she sat up. \" i can\\'t believe it\\'s real. i\\'ve been dreaming about you for the last two days. \" \\n \" i can\\'t believe it either. i guess that\\'s how'}]"
|
||||
]
|
||||
},
|
||||
"execution_count": 8,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"prompt = \"It was early evening when I can back from work. I usually work late, but this time it was an exception. When I entered a room, I saw\"\n",
|
||||
"generator(prompt,max_length=100,do_sample=True,temperature=0.8)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們還可以為採樣提供額外的參數： \n",
|
||||
"* `top_k` 指定在使用採樣時需要考慮的詞彙選項數量。這可以減少在文本中出現奇怪（低概率）詞彙的機會。 \n",
|
||||
"* `top_p` 類似，但我們選擇的是總概率大於 p 的最可能詞彙的最小子集。 \n",
|
||||
"\n",
|
||||
"可以隨意嘗試加入這些參數進行實驗。 \n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"attachments": {},
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 微調您的模型\n",
|
||||
"\n",
|
||||
"您亦可使用自己的數據集[微調您的模型](https://learn.microsoft.com/en-us/azure/cognitive-services/openai/how-to/fine-tuning?pivots=programming-language-studio?WT.mc_id=academic-77998-bethanycheum)。這樣可以在保留語言模型主要部分的同時，調整文本的風格。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件已使用人工智能翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。應以原文文件作為權威來源。對於關鍵資訊，建議尋求專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或誤釋不承擔責任。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"interpreter": {
|
||||
"hash": "16af2a8bbb083ea23e5e41c7f5787656b2ce26968575d8763f2c4b17f9cd711f"
|
||||
},
|
||||
"kernelspec": {
|
||||
"display_name": "Python 3.8.12 ('py38')",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.10.11"
|
||||
},
|
||||
"orig_nbformat": 4,
|
||||
"coopTranslator": {
|
||||
"original_hash": "d4ff89615d38924a55594f16d6d20678",
|
||||
"translation_date": "2025-08-31T10:44:25+00:00",
|
||||
"source_file": "lessons/5-NLP/20-LangModels/GPT-PyTorch.ipynb",
|
||||
"language_code": "hk"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 2
|
||||
}
|
||||
|
|
@ -0,0 +1,49 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 作業：丟番圖方程\n",
|
||||
"\n",
|
||||
"> 此作業是 [AI for Beginners Curriculum](http://github.com/microsoft/ai-for-beginners) 的一部分，靈感來自 [這篇文章](https://habr.com/post/128704/)。\n",
|
||||
"\n",
|
||||
"你的目標是解決所謂的 **丟番圖方程**——一種具有整數根和整數係數的方程。例如，考慮以下方程：\n",
|
||||
"\n",
|
||||
"$$a+2b+3c+4d=30$$\n",
|
||||
"\n",
|
||||
"你需要找到整數根 $a$,$b$,$c$,$d\\in\\mathbb{N}$，使其滿足這個方程。\n",
|
||||
"\n",
|
||||
"提示：\n",
|
||||
"1. 你可以考慮根值在區間 [0;30] 之內\n",
|
||||
"1. 作為基因，可以考慮使用根值的列表\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": []
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件已使用人工智能翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。儘管我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。原始文件的母語版本應被視為權威來源。對於重要信息，建議使用專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或錯誤解釋不承擔責任。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"language_info": {
|
||||
"name": "python"
|
||||
},
|
||||
"orig_nbformat": 4,
|
||||
"coopTranslator": {
|
||||
"original_hash": "a967e1fa1e11ab2b6467b19349a4a9aa",
|
||||
"translation_date": "2025-08-31T09:21:19+00:00",
|
||||
"source_file": "lessons/6-Other/21-GeneticAlgorithms/Diophantine.ipynb",
|
||||
"language_code": "hk"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 2
|
||||
}
|
||||
File diff suppressed because one or more lines are too long
|
|
@ -0,0 +1,501 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 使用強化學習訓練模型平衡Cartpole\n",
|
||||
"\n",
|
||||
"這份筆記本是 [AI for Beginners Curriculum](http://aka.ms/ai-beginners) 的一部分，靈感來自 [官方 PyTorch 教程](https://pytorch.org/tutorials/intermediate/reinforcement_q_learning.html) 和 [這個 Cartpole PyTorch 實現](https://github.com/yc930401/Actor-Critic-pytorch)。\n",
|
||||
"\n",
|
||||
"在這個例子中，我們將使用強化學習（RL）訓練一個模型，使其能夠平衡一根安裝在小車上的杆。小車可以在水平軸上向左或向右移動。我們將使用 [OpenAI Gym](https://www.gymlibrary.ml/) 環境來模擬這個場景。\n",
|
||||
"\n",
|
||||
"> **注意**：你可以在本地（例如使用 Visual Studio Code）運行這節課的代碼，此時模擬器會在新窗口中打開。如果在線運行代碼，你可能需要對代碼進行一些調整，具體請參考 [這裡](https://towardsdatascience.com/rendering-openai-gym-envs-on-binder-and-google-colab-536f99391cc7)。\n",
|
||||
"\n",
|
||||
"我們將從確保 Gym 已安裝開始：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import sys\n",
|
||||
"!{sys.executable} -m pip install gym"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在讓我們建立 CartPole 環境，並了解如何操作它。一個環境具有以下特性：\n",
|
||||
"\n",
|
||||
"* **動作空間** 是我們在模擬的每一步中可以執行的所有可能動作的集合 \n",
|
||||
"* **觀察空間** 是我們可以進行觀察的範圍\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import gym\n",
|
||||
"\n",
|
||||
"env = gym.make(\"CartPole-v1\")\n",
|
||||
"\n",
|
||||
"print(f\"Action space: {env.action_space}\")\n",
|
||||
"print(f\"Observation space: {env.observation_space}\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"讓我們看看模擬是如何運作的。以下的迴圈會執行模擬，直到 `env.step` 不再返回終止標誌 `done` 為止。我們會使用 `env.action_space.sample()` 隨機選擇動作，這意味著實驗可能會非常快地失敗（當小車的速度、位置或角度超出某些限制時，CartPole 環境就會終止）。\n",
|
||||
"\n",
|
||||
"> 模擬將會在新視窗中開啟。你可以多次執行程式碼，觀察它的行為。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"env.reset()\n",
|
||||
"\n",
|
||||
"done = False\n",
|
||||
"total_reward = 0\n",
|
||||
"while not done:\n",
|
||||
" env.render()\n",
|
||||
" obs, rew, done, info = env.step(env.action_space.sample())\n",
|
||||
" total_reward += rew\n",
|
||||
" print(f\"{obs} -> {rew}\")\n",
|
||||
"print(f\"Total reward: {total_reward}\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"你可以注意到觀察值包含四個數字，它們分別是：\n",
|
||||
"- 小車的位置\n",
|
||||
"- 小車的速度\n",
|
||||
"- 杆子的角度\n",
|
||||
"- 杆子的旋轉速率\n",
|
||||
"\n",
|
||||
"`rew` 是我們在每一步獲得的獎勵。在 CartPole 環境中，每模擬一步你都會獲得 1 分，目標是最大化總獎勵，也就是讓 CartPole 能夠平衡而不倒下的時間越長越好。\n",
|
||||
"\n",
|
||||
"在強化學習中，我們的目標是訓練一個 **策略** $\\pi$，它會根據每個狀態 $s$ 告訴我們應該採取的行動 $a$，基本上就是 $a = \\pi(s)$。\n",
|
||||
"\n",
|
||||
"如果你想要一個概率性的解法，可以將策略理解為對每個行動返回一組概率，也就是 $\\pi(a|s)$ 表示在狀態 $s$ 下採取行動 $a$ 的概率。\n",
|
||||
"\n",
|
||||
"## 策略梯度方法\n",
|
||||
"\n",
|
||||
"在最簡單的強化學習算法中，稱為 **策略梯度**，我們將訓練一個神經網絡來預測下一步的行動。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import numpy as np\n",
|
||||
"import matplotlib.pyplot as plt\n",
|
||||
"import torch\n",
|
||||
"\n",
|
||||
"num_inputs = 4\n",
|
||||
"num_actions = 2\n",
|
||||
"\n",
|
||||
"model = torch.nn.Sequential(\n",
|
||||
" torch.nn.Linear(num_inputs, 128, bias=False, dtype=torch.float32),\n",
|
||||
" torch.nn.ReLU(),\n",
|
||||
" torch.nn.Linear(128, num_actions, bias = False, dtype=torch.float32),\n",
|
||||
" torch.nn.Softmax(dim=1)\n",
|
||||
")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們將通過進行多次實驗來訓練網絡，並在每次運行後更新網絡。讓我們定義一個函數來運行實驗並返回結果（即所謂的**追蹤**）——所有狀態、行動（及其建議的概率）和獎勵：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"def run_episode(max_steps_per_episode = 10000,render=False): \n",
|
||||
" states, actions, probs, rewards = [],[],[],[]\n",
|
||||
" state = env.reset()\n",
|
||||
" for _ in range(max_steps_per_episode):\n",
|
||||
" if render:\n",
|
||||
" env.render()\n",
|
||||
" action_probs = model(torch.from_numpy(np.expand_dims(state,0)))[0]\n",
|
||||
" action = np.random.choice(num_actions, p=np.squeeze(action_probs.detach().numpy()))\n",
|
||||
" nstate, reward, done, info = env.step(action)\n",
|
||||
" if done:\n",
|
||||
" break\n",
|
||||
" states.append(state)\n",
|
||||
" actions.append(action)\n",
|
||||
" probs.append(action_probs.detach().numpy())\n",
|
||||
" rewards.append(reward)\n",
|
||||
" state = nstate\n",
|
||||
" return np.vstack(states), np.vstack(actions), np.vstack(probs), np.vstack(rewards)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"您可以使用未訓練的網絡運行一個劇集，並觀察到總回報（即劇集的長度）非常低：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"s, a, p, r = run_episode()\n",
|
||||
"print(f\"Total reward: {np.sum(r)}\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"政策梯度算法的一個棘手方面是使用**折扣回報**。其想法是我們在遊戲的每一步計算總回報的向量，並在此過程中使用某個係數 $gamma$ 對早期回報進行折扣。我們還會對結果向量進行標準化，因為我們將使用它作為權重來影響我們的訓練：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"eps = 0.0001\n",
|
||||
"\n",
|
||||
"def discounted_rewards(rewards,gamma=0.99,normalize=True):\n",
|
||||
" ret = []\n",
|
||||
" s = 0\n",
|
||||
" for r in rewards[::-1]:\n",
|
||||
" s = r + gamma * s\n",
|
||||
" ret.insert(0, s)\n",
|
||||
" if normalize:\n",
|
||||
" ret = (ret-np.mean(ret))/(np.std(ret)+eps)\n",
|
||||
" return ret"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在讓我們開始實際的訓練！我們將運行 300 次實驗，每次實驗中我們將執行以下步驟：\n",
|
||||
"\n",
|
||||
"1. 運行實驗並收集追蹤數據\n",
|
||||
"2. 計算所採取行動與預測機率之間的差異（`gradients`）。差異越小，表示我們越確定採取了正確的行動。\n",
|
||||
"3. 計算折扣後的回報，並將梯度乘以折扣後的回報——這樣可以確保高回報的步驟對最終結果的影響比低回報的步驟更大。\n",
|
||||
"4. 我們的神經網絡的預期目標行動部分來自運行期間的預測機率，部分來自計算出的梯度。我們將使用 `alpha` 參數來決定梯度和回報在多大程度上被考慮——這被稱為強化學習算法的*學習率*。\n",
|
||||
"5. 最後，我們基於狀態和預期行動訓練我們的網絡，並重複這個過程。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"optimizer = torch.optim.Adam(model.parameters(), lr=0.01)\n",
|
||||
"\n",
|
||||
"def train_on_batch(x, y):\n",
|
||||
" x = torch.from_numpy(x)\n",
|
||||
" y = torch.from_numpy(y)\n",
|
||||
" optimizer.zero_grad()\n",
|
||||
" predictions = model(x)\n",
|
||||
" loss = -torch.mean(torch.log(predictions) * y)\n",
|
||||
" loss.backward()\n",
|
||||
" optimizer.step()\n",
|
||||
" return loss"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"alpha = 1e-4\n",
|
||||
"\n",
|
||||
"history = []\n",
|
||||
"for epoch in range(300):\n",
|
||||
" states, actions, probs, rewards = run_episode()\n",
|
||||
" one_hot_actions = np.eye(2)[actions.T][0]\n",
|
||||
" gradients = one_hot_actions-probs\n",
|
||||
" dr = discounted_rewards(rewards)\n",
|
||||
" gradients *= dr\n",
|
||||
" target = alpha*np.vstack([gradients])+probs\n",
|
||||
" train_on_batch(states,target)\n",
|
||||
" history.append(np.sum(rewards))\n",
|
||||
" if epoch%100==0:\n",
|
||||
" print(f\"{epoch} -> {np.sum(rewards)}\")\n",
|
||||
"\n",
|
||||
"plt.plot(history)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在讓我們運行帶有渲染的劇集來查看結果：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"_ = run_episode(render=True)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"希望你可以看到，現在杆子已經能夠相當穩定地平衡了！\n",
|
||||
"\n",
|
||||
"## Actor-Critic 模型\n",
|
||||
"\n",
|
||||
"Actor-Critic 模型是策略梯度的進一步發展，在這個模型中，我們建立了一個神經網絡來同時學習策略和估算的回報。這個網絡會有兩個輸出（或者你可以將其視為兩個獨立的網絡）：\n",
|
||||
"* **Actor** 會根據策略梯度模型，通過提供狀態概率分佈來推薦應採取的行動。\n",
|
||||
"* **Critic** 則會估算這些行動可能帶來的回報。它會在給定的狀態下返回未來的總估算回報。\n",
|
||||
"\n",
|
||||
"讓我們定義這樣一個模型：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"from itertools import count\n",
|
||||
"import torch.nn.functional as F"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n",
|
||||
"env = gym.make(\"CartPole-v1\")\n",
|
||||
"\n",
|
||||
"state_size = env.observation_space.shape[0]\n",
|
||||
"action_size = env.action_space.n\n",
|
||||
"lr = 0.0001\n",
|
||||
"\n",
|
||||
"class Actor(torch.nn.Module):\n",
|
||||
" def __init__(self, state_size, action_size):\n",
|
||||
" super(Actor, self).__init__()\n",
|
||||
" self.state_size = state_size\n",
|
||||
" self.action_size = action_size\n",
|
||||
" self.linear1 = torch.nn.Linear(self.state_size, 128)\n",
|
||||
" self.linear2 = torch.nn.Linear(128, 256)\n",
|
||||
" self.linear3 = torch.nn.Linear(256, self.action_size)\n",
|
||||
"\n",
|
||||
" def forward(self, state):\n",
|
||||
" output = F.relu(self.linear1(state))\n",
|
||||
" output = F.relu(self.linear2(output))\n",
|
||||
" output = self.linear3(output)\n",
|
||||
" distribution = torch.distributions.Categorical(F.softmax(output, dim=-1))\n",
|
||||
" return distribution\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"class Critic(torch.nn.Module):\n",
|
||||
" def __init__(self, state_size, action_size):\n",
|
||||
" super(Critic, self).__init__()\n",
|
||||
" self.state_size = state_size\n",
|
||||
" self.action_size = action_size\n",
|
||||
" self.linear1 = torch.nn.Linear(self.state_size, 128)\n",
|
||||
" self.linear2 = torch.nn.Linear(128, 256)\n",
|
||||
" self.linear3 = torch.nn.Linear(256, 1)\n",
|
||||
"\n",
|
||||
" def forward(self, state):\n",
|
||||
" output = F.relu(self.linear1(state))\n",
|
||||
" output = F.relu(self.linear2(output))\n",
|
||||
" value = self.linear3(output)\n",
|
||||
" return value"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們需要稍微修改我們的 `discounted_rewards` 和 `run_episode` 函數：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"def discounted_rewards(next_value, rewards, masks, gamma=0.99):\n",
|
||||
" R = next_value\n",
|
||||
" returns = []\n",
|
||||
" for step in reversed(range(len(rewards))):\n",
|
||||
" R = rewards[step] + gamma * R * masks[step]\n",
|
||||
" returns.insert(0, R)\n",
|
||||
" return returns\n",
|
||||
"\n",
|
||||
"def run_episode(actor, critic, n_iters):\n",
|
||||
" optimizerA = torch.optim.Adam(actor.parameters())\n",
|
||||
" optimizerC = torch.optim.Adam(critic.parameters())\n",
|
||||
" for iter in range(n_iters):\n",
|
||||
" state = env.reset()\n",
|
||||
" log_probs = []\n",
|
||||
" values = []\n",
|
||||
" rewards = []\n",
|
||||
" masks = []\n",
|
||||
" entropy = 0\n",
|
||||
" env.reset()\n",
|
||||
"\n",
|
||||
" for i in count():\n",
|
||||
" env.render()\n",
|
||||
" state = torch.FloatTensor(state).to(device)\n",
|
||||
" dist, value = actor(state), critic(state)\n",
|
||||
"\n",
|
||||
" action = dist.sample()\n",
|
||||
" next_state, reward, done, _ = env.step(action.cpu().numpy())\n",
|
||||
"\n",
|
||||
" log_prob = dist.log_prob(action).unsqueeze(0)\n",
|
||||
" entropy += dist.entropy().mean()\n",
|
||||
"\n",
|
||||
" log_probs.append(log_prob)\n",
|
||||
" values.append(value)\n",
|
||||
" rewards.append(torch.tensor([reward], dtype=torch.float, device=device))\n",
|
||||
" masks.append(torch.tensor([1-done], dtype=torch.float, device=device))\n",
|
||||
"\n",
|
||||
" state = next_state\n",
|
||||
"\n",
|
||||
" if done:\n",
|
||||
" print('Iteration: {}, Score: {}'.format(iter, i))\n",
|
||||
" break\n",
|
||||
"\n",
|
||||
"\n",
|
||||
" next_state = torch.FloatTensor(next_state).to(device)\n",
|
||||
" next_value = critic(next_state)\n",
|
||||
" returns = discounted_rewards(next_value, rewards, masks)\n",
|
||||
"\n",
|
||||
" log_probs = torch.cat(log_probs)\n",
|
||||
" returns = torch.cat(returns).detach()\n",
|
||||
" values = torch.cat(values)\n",
|
||||
"\n",
|
||||
" advantage = returns - values\n",
|
||||
"\n",
|
||||
" actor_loss = -(log_probs * advantage.detach()).mean()\n",
|
||||
" critic_loss = advantage.pow(2).mean()\n",
|
||||
"\n",
|
||||
" optimizerA.zero_grad()\n",
|
||||
" optimizerC.zero_grad()\n",
|
||||
" actor_loss.backward()\n",
|
||||
" critic_loss.backward()\n",
|
||||
" optimizerA.step()\n",
|
||||
" optimizerC.step()\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在我們將運行主要訓練循環。我們將通過計算適當的損失函數並更新網絡參數來使用手動網絡訓練過程：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"\n",
|
||||
"actor = Actor(state_size, action_size).to(device)\n",
|
||||
"critic = Critic(state_size, action_size).to(device)\n",
|
||||
"run_episode(actor, critic, n_iters=100)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": []
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"env.close()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 重點\n",
|
||||
"\n",
|
||||
"在這個示範中，我們看到了兩種強化學習算法：簡單的策略梯度和更複雜的演員-評論家方法。你可以看到這些算法是基於狀態、行動和獎勵的抽象概念運作的，因此它們可以應用於非常不同的環境。\n",
|
||||
"\n",
|
||||
"強化學習讓我們只需通過觀察最終的獎勵，就能學習解決問題的最佳策略。不需要標籤數據集的特性，讓我們可以多次重複模擬來優化模型。然而，強化學習仍然面臨許多挑戰，如果你決定更深入研究這個有趣的人工智能領域，可能會學到更多相關知識。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件已使用人工智能翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。儘管我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。原始語言的文件應被視為權威來源。對於重要資訊，建議使用專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或錯誤解釋概不負責。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"kernelspec": {
|
||||
"display_name": "Python 3.10.4 64-bit",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.10.4"
|
||||
},
|
||||
"orig_nbformat": 4,
|
||||
"vscode": {
|
||||
"interpreter": {
|
||||
"hash": "916dbcbb3f70747c44a77c7bcd40155683ae19c65e1c03b4aa3499c5328201f1"
|
||||
}
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "04f8d9978cd11281d81dd037cbf6ce20",
|
||||
"translation_date": "2025-08-31T09:26:45+00:00",
|
||||
"source_file": "lessons/6-Other/22-DeepRL/CartPole-RL-PyTorch.ipynb",
|
||||
"language_code": "hk"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 2
|
||||
}
|
||||
File diff suppressed because one or more lines are too long
|
|
@ -0,0 +1,109 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 訓練山地車逃脫\n",
|
||||
"\n",
|
||||
"來自 [AI for Beginners Curriculum](https://github.com/microsoft/ai-for-beginners) 的實驗作業。\n",
|
||||
"\n",
|
||||
"你的目標是訓練強化學習代理控制 OpenAI 環境中的 [Mountain Car](https://www.gymlibrary.ml/environments/classic_control/mountain_car/)。\n",
|
||||
"\n",
|
||||
"讓我們從創建環境開始：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 1,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import gym\n",
|
||||
"env = gym.make('MountainCar-v0')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"讓我們看看隨機實驗是什麼樣子：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 2,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"state = env.reset()\n",
|
||||
"while True:\n",
|
||||
" env.render()\n",
|
||||
" action = env.action_space.sample()\n",
|
||||
" state, reward, done, info = env.step(action)\n",
|
||||
" if done:\n",
|
||||
" break"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": []
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"## Lost of code here"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"env.close()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件已使用人工智能翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。儘管我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。原始文件的母語版本應被視為權威來源。對於重要信息，建議使用專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或錯誤解釋概不負責。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"interpreter": {
|
||||
"hash": "16af2a8bbb083ea23e5e41c7f5787656b2ce26968575d8763f2c4b17f9cd711f"
|
||||
},
|
||||
"kernelspec": {
|
||||
"display_name": "Python 3.8.12 ('py38')",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.8.12"
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "f062b3b18449593ef8e0fcc029868781",
|
||||
"translation_date": "2025-08-31T09:29:01+00:00",
|
||||
"source_file": "lessons/6-Other/22-DeepRL/lab/MountainCar.ipynb",
|
||||
"language_code": "hk"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 4
|
||||
}
|
||||
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
|
|
@ -1,8 +1,8 @@
|
|||
<!--
|
||||
CO_OP_TRANSLATOR_METADATA:
|
||||
{
|
||||
"original_hash": "f3a6b0ddf7e6e3f33b2a543baf086dc9",
|
||||
"translation_date": "2025-08-24T21:39:26+00:00",
|
||||
"original_hash": "07191303b7ea2aff1d47e2b0fe4bb862",
|
||||
"translation_date": "2025-08-31T09:16:49+00:00",
|
||||
"source_file": "README.md",
|
||||
"language_code": "tw"
|
||||
}
|
||||
|
|
@ -21,107 +21,119 @@ CO_OP_TRANSLATOR_METADATA:
|
|||
|
||||
[](https://discord.gg/zxKYvhSnVp?WT.mc_id=academic-000002-leestott)
|
||||
|
||||
# 人工智慧入門課程 - 課程大綱
|
||||
# 人工智慧入門課程 - 一個完整的課程
|
||||
|
||||
| 繪製的速寫筆記](./lessons/sketchnotes/ai-overview.png)|
|
||||
|:---:|
|
||||
| 人工智慧入門課程 - _速寫筆記由 [@girlie_mac](https://twitter.com/girlie_mac)_ |
|
||||
||
|
||||
|:---:|
|
||||
| 人工智慧入門課程 - _速記圖由 [@girlie_mac](https://twitter.com/girlie_mac) 繪製_ |
|
||||
|
||||
探索 **人工智慧** (AI) 的世界，透過我們為期 12 週、共 24 節課的課程！課程包含實作教學、測驗及實驗室練習。這是一個適合初學者的課程，涵蓋 TensorFlow 和 PyTorch 等工具，以及人工智慧的倫理議題。
|
||||
透過我們為期 12 週、共 24 節課的課程，探索 **人工智慧** (AI) 的世界！課程包含實用的教學、測驗和實驗室練習。這是一個適合初學者的課程，涵蓋了像 TensorFlow 和 PyTorch 這樣的工具，以及 AI 的倫理議題。
|
||||
|
||||
## 您將學到什麼
|
||||
### 🌐 多語言支援
|
||||
|
||||
**[課程心智圖](http://soshnikov.com/courses/ai-for-beginners/mindmap.html)**
|
||||
#### 透過 GitHub Action 支援（自動化且始終保持最新）
|
||||
|
||||
在這門課程中，您將學到：
|
||||
[法文](../fr/README.md) | [西班牙文](../es/README.md) | [德文](../de/README.md) | [俄文](../ru/README.md) | [阿拉伯文](../ar/README.md) | [波斯文 (法爾西)](../fa/README.md) | [烏爾都文](../ur/README.md) | [中文 (簡體)](../zh/README.md) | [中文 (繁體，澳門)](../mo/README.md) | [中文 (繁體，香港)](../hk/README.md) | [中文 (繁體，台灣)](./README.md) | [日文](../ja/README.md) | [韓文](../ko/README.md) | [印地文](../hi/README.md) | [孟加拉文](../bn/README.md) | [馬拉地文](../mr/README.md) | [尼泊爾文](../ne/README.md) | [旁遮普文 (古木基文)](../pa/README.md) | [葡萄牙文 (葡萄牙)](../pt/README.md) | [葡萄牙文 (巴西)](../br/README.md) | [義大利文](../it/README.md) | [波蘭文](../pl/README.md) | [土耳其文](../tr/README.md) | [希臘文](../el/README.md) | [泰文](../th/README.md) | [瑞典文](../sv/README.md) | [丹麥文](../da/README.md) | [挪威文](../no/README.md) | [芬蘭文](../fi/README.md) | [荷蘭文](../nl/README.md) | [希伯來文](../he/README.md) | [越南文](../vi/README.md) | [印尼文](../id/README.md) | [馬來文](../ms/README.md) | [他加祿文 (菲律賓)](../tl/README.md) | [斯瓦希里文](../sw/README.md) | [匈牙利文](../hu/README.md) | [捷克文](../cs/README.md) | [斯洛伐克文](../sk/README.md) | [羅馬尼亞文](../ro/README.md) | [保加利亞文](../bg/README.md) | [塞爾維亞文 (西里爾字母)](../sr/README.md) | [克羅埃西亞文](../hr/README.md) | [斯洛維尼亞文](../sl/README.md) | [烏克蘭文](../uk/README.md) | [緬甸文 (緬甸)](../my/README.md)
|
||||
|
||||
* 不同的人工智慧方法，包括「經典」的符號方法，透過 **知識表示** 和推理 ([GOFAI](https://en.wikipedia.org/wiki/Symbolic_artificial_intelligence))。
|
||||
* **神經網路** 和 **深度學習**，這是現代人工智慧的核心。我們將使用兩個最受歡迎的框架 [TensorFlow](http://Tensorflow.org) 和 [PyTorch](http://pytorch.org) 的程式碼來說明這些重要主題背後的概念。
|
||||
* 用於處理影像和文字的 **神經架構**。我們將涵蓋近期的模型，但可能不包括最新的技術。
|
||||
* 不太常見的人工智慧方法，例如 **遺傳演算法** 和 **多代理系統**。
|
||||
**如果您希望支援其他語言，請參考 [這裡](https://github.com/Azure/co-op-translator/blob/main/getting_started/supported-languages.md)**
|
||||
|
||||
我們不會在這門課程中涵蓋：
|
||||
## 加入社群
|
||||
[](https://discord.gg/kzRShWzttr)
|
||||
|
||||
> [在 Microsoft Learn 集合中找到所有額外的課程資源](https://learn.microsoft.com/en-us/collections/7w28iy2xrqzdj0?WT.mc_id=academic-77998-bethanycheum)
|
||||
## 您將學到什麼
|
||||
|
||||
* **人工智慧在商業中的應用案例**。建議參加 [商業用戶的人工智慧入門](https://docs.microsoft.com/learn/paths/introduction-ai-for-business-users/?WT.mc_id=academic-77998-bethanycheum) 學習路徑，或由 [INSEAD](https://www.insead.edu/) 合作開發的 [AI 商業學院](https://www.microsoft.com/ai/ai-business-school/?WT.mc_id=academic-77998-bethanycheum)。
|
||||
* **經典機器學習**，這在我們的 [機器學習入門課程](http://github.com/Microsoft/ML-for-Beginners) 中有詳細描述。
|
||||
* 使用 **[認知服務](https://azure.microsoft.com/services/cognitive-services/?WT.mc_id=academic-77998-bethanycheum)** 建立的人工智慧應用。建議從 Microsoft Learn 的模組開始，例如 [視覺](https://docs.microsoft.com/learn/paths/create-computer-vision-solutions-azure-cognitive-services/?WT.mc_id=academic-77998-bethanycheum)、[自然語言處理](https://docs.microsoft.com/learn/paths/explore-natural-language-processing/?WT.mc_id=academic-77998-bethanycheum)、**[Azure OpenAI Service 的生成式人工智慧](https://learn.microsoft.com/en-us/training/paths/develop-ai-solutions-azure-openai/?WT.mc_id=academic-77998-bethanycheum)** 等。
|
||||
* 特定的機器學習 **雲端框架**，例如 [Azure Machine Learning](https://azure.microsoft.com/services/machine-learning/?WT.mc_id=academic-77998-bethanycheum)、[Microsoft Fabric](https://learn.microsoft.com/en-us/training/paths/get-started-fabric/?WT.mc_id=academic-77998-bethanycheum) 或 [Azure Databricks](https://docs.microsoft.com/learn/paths/data-engineer-azure-databricks?WT.mc_id=academic-77998-bethanycheum)。建議參加 [使用 Azure Machine Learning 建立和操作機器學習解決方案](https://docs.microsoft.com/learn/paths/build-ai-solutions-with-azure-ml-service/?WT.mc_id=academic-77998-bethanycheum) 和 [使用 Azure Databricks 建立和操作機器學習解決方案](https://docs.microsoft.com/learn/paths/build-operate-machine-learning-solutions-azure-databricks/?WT.mc_id=academic-77998-bethanycheum) 學習路徑。
|
||||
* **對話式人工智慧** 和 **聊天機器人**。有一個單獨的 [建立對話式人工智慧解決方案](https://docs.microsoft.com/learn/paths/create-conversational-ai-solutions/?WT.mc_id=academic-77998-bethanycheum) 學習路徑，您也可以參考 [這篇部落格文章](https://soshnikov.com/azure/hello-bot-conversational-ai-on-microsoft-platform/) 了解更多細節。
|
||||
* **深度學習的數學基礎**。建議參考 Ian Goodfellow、Yoshua Bengio 和 Aaron Courville 所著的 [深度學習](https://www.amazon.com/Deep-Learning-Adaptive-Computation-Machine/dp/0262035618)，也可在線上取得：[https://www.deeplearningbook.org/](https://www.deeplearningbook.org/)。
|
||||
**[課程心智圖](http://soshnikov.com/courses/ai-for-beginners/mindmap.html)**
|
||||
|
||||
若想輕鬆入門 _雲端人工智慧_ 主題，建議參加 [在 Azure 上開始人工智慧學習](https://docs.microsoft.com/learn/paths/get-started-with-artificial-intelligence-on-azure/?WT.mc_id=academic-77998-bethanycheum) 學習路徑。
|
||||
在這個課程中，您將學到：
|
||||
|
||||
# 課程內容
|
||||
* 不同的人工智慧方法，包括「經典的」符號方法，使用 **知識表徵** 和推理（[GOFAI](https://en.wikipedia.org/wiki/Symbolic_artificial_intelligence)）。
|
||||
* **神經網路** 和 **深度學習**，這是現代 AI 的核心。我們將使用兩個最受歡迎的框架 [TensorFlow](http://Tensorflow.org) 和 [PyTorch](http://pytorch.org) 的程式碼來說明這些重要主題的概念。
|
||||
* 用於處理影像和文字的 **神經架構**。我們將涵蓋一些近期的模型，但可能不包括最前沿的技術。
|
||||
* 不太常見的 AI 方法，例如 **遺傳演算法** 和 **多代理系統**。
|
||||
|
||||
| | 課程連結 | PyTorch/Keras/TensorFlow | 實驗室 |
|
||||
| :-: | :------------------------------------------------------------------------------------------------------------------------------------------: | :---------------------------------------------------------------------------------------------: | ------------------------------------------------------------------------------ |
|
||||
| 0 | [課程設定](./lessons/0-course-setup/setup.md) | [設定您的開發環境](./lessons/0-course-setup/how-to-run.md) | |
|
||||
| I | [**人工智慧簡介**](./lessons/1-Intro/README.md) | | |
|
||||
| 01 | [人工智慧的簡介與歷史](./lessons/1-Intro/README.md) | - | - |
|
||||
| II | **符號人工智慧** |
|
||||
| 02 | [知識表示與專家系統](./lessons/2-Symbolic/README.md) | [專家系統](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/2-Symbolic/Animals.ipynb) / [本體論](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/2-Symbolic/FamilyOntology.ipynb) /[概念圖](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/2-Symbolic/MSConceptGraph.ipynb) | |
|
||||
| III | [**神經網路簡介**](./lessons/3-NeuralNetworks/README.md) |||
|
||||
| 03 | [感知器](./lessons/3-NeuralNetworks/03-Perceptron/README.md) | [筆記本](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/3-NeuralNetworks/03-Perceptron/Perceptron.ipynb) | [實驗室](./lessons/3-NeuralNetworks/03-Perceptron/lab/README.md) |
|
||||
| 04 | [多層感知器與建立我們自己的框架](./lessons/3-NeuralNetworks/04-OwnFramework/README.md) | [筆記本](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/3-NeuralNetworks/04-OwnFramework/OwnFramework.ipynb) | [實驗室](./lessons/3-NeuralNetworks/04-OwnFramework/lab/README.md) |
|
||||
| 05 | [框架簡介 (PyTorch/TensorFlow) 與過度擬合](./lessons/3-NeuralNetworks/05-Frameworks/README.md) | [PyTorch](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/3-NeuralNetworks/05-Frameworks/IntroPyTorch.ipynb) / [Keras](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/3-NeuralNetworks/05-Frameworks/IntroKeras.ipynb) / [TensorFlow](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/3-NeuralNetworks/05-Frameworks/IntroKerasTF.ipynb) | [實驗室](./lessons/3-NeuralNetworks/05-Frameworks/lab/README.md) |
|
||||
| IV | [**電腦視覺**](./lessons/4-ComputerVision/README.md) | [PyTorch](https://docs.microsoft.com/learn/modules/intro-computer-vision-pytorch/?WT.mc_id=academic-77998-cacaste) / [TensorFlow](https://docs.microsoft.com/learn/modules/intro-computer-vision-TensorFlow/?WT.mc_id=academic-77998-cacaste)| [探索 Microsoft Azure 上的電腦視覺](https://learn.microsoft.com/en-us/collections/7w28iy2xrqzdj0?WT.mc_id=academic-77998-bethanycheum) |
|
||||
| 06 | [電腦視覺簡介. OpenCV](./lessons/4-ComputerVision/06-IntroCV/README.md) | [筆記本](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/4-ComputerVision/06-IntroCV/OpenCV.ipynb) | [實驗室](./lessons/4-ComputerVision/06-IntroCV/lab/README.md) |
|
||||
| 07 | [卷積神經網路](./lessons/4-ComputerVision/07-ConvNets/README.md) & [CNN 架構](./lessons/4-ComputerVision/07-ConvNets/CNN_Architectures.md) | [PyTorch](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/4-ComputerVision/07-ConvNets/ConvNetsPyTorch.ipynb) /[TensorFlow](https://microsoft.github.io/AI-For-Beginners/lessons/4-ComputerVision/07-ConvNets/ConvNetsTF.ipynb) | [實驗室](./lessons/4-ComputerVision/07-ConvNets/lab/README.md) |
|
||||
| 08 | [預訓練網路與遷移學習](./lessons/4-ComputerVision/08-TransferLearning/README.md) 和 [訓練技巧](./lessons/4-ComputerVision/08-TransferLearning/TrainingTricks.md) | [PyTorch](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/4-ComputerVision/08-TransferLearning/TransferLearningPyTorch.ipynb) / [TensorFlow](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/3-NeuralNetworks/05-Frameworks/IntroKerasTF.ipynb) | [實驗室](./lessons/4-ComputerVision/08-TransferLearning/lab/README.md) |
|
||||
| 09 | [自編碼器與變分自編碼器 (VAEs)](./lessons/4-ComputerVision/09-Autoencoders/README.md) | [PyTorch](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/4-ComputerVision/09-Autoencoders/AutoEncodersPyTorch.ipynb) / [TensorFlow](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/4-ComputerVision/09-Autoencoders/AutoencodersTF.ipynb) | |
|
||||
| 10 | [生成對抗網路 (GANs) 與藝術風格轉換](./lessons/4-ComputerVision/10-GANs/README.md) | [PyTorch](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/4-ComputerVision/10-GANs/GANPyTorch.ipynb) / [TensorFlow](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/4-ComputerVision/10-GANs/GANTF.ipynb) | |
|
||||
| 11 | [物件偵測](./lessons/4-ComputerVision/11-ObjectDetection/README.md) | [TensorFlow](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/4-ComputerVision/11-ObjectDetection/ObjectDetection.ipynb) | [實驗室](./lessons/4-ComputerVision/11-ObjectDetection/lab/README.md) |
|
||||
| 12 | [語義分割. U-Net](./lessons/4-ComputerVision/12-Segmentation/README.md) | [PyTorch](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/4-ComputerVision/12-Segmentation/SemanticSegmentationPytorch.ipynb) / [TensorFlow](../../(https:/github.com/microsoft/AI-For-Beginners/blob/main/lessons/4-ComputerVision/12-Segmentation/SemanticSegmentationTF.ipynb)) | |
|
||||
| V | [**自然語言處理**](./lessons/5-NLP/README.md) | [PyTorch](https://docs.microsoft.com/learn/modules/intro-natural-language-processing-pytorch/?WT.mc_id=academic-77998-cacaste) /[TensorFlow](https://docs.microsoft.com/learn/modules/intro-natural-language-processing-TensorFlow/?WT.mc_id=academic-77998-cacaste) | [在 Microsoft Azure 上探索自然語言處理](https://learn.microsoft.com/en-us/collections/7w28iy2xrqzdj0?WT.mc_id=academic-77998-bethanycheum)|
|
||||
| 13 | [文本表示. Bow/TF-IDF](./lessons/5-NLP/13-TextRep/README.md) | [PyTorch](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/5-NLP/13-TextRep/TextRepresentationPyTorch.ipynb) / [TensorFlow](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/5-NLP/13-TextRep/TextRepresentationTF.ipynb) | |
|
||||
| 14 | [語義詞嵌入. Word2Vec 和 GloVe](./lessons/5-NLP/14-Embeddings/README.md) | [PyTorch](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/5-NLP/14-Embeddings/EmbeddingsPyTorch.ipynb) / [TensorFlow](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/5-NLP/14-Embeddings/EmbeddingsTF.ipynb) | |
|
||||
| 15 | [語言建模. 訓練自己的嵌入](./lessons/5-NLP/15-LanguageModeling/README.md) | [PyTorch](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/5-NLP/15-LanguageModeling/CBoW-PyTorch.ipynb) / [TensorFlow](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/5-NLP/15-LanguageModeling/CBoW-TF.ipynb) | [實驗室](./lessons/5-NLP/15-LanguageModeling/lab/README.md) |
|
||||
| 16 | [循環神經網路 (RNN)](./lessons/5-NLP/16-RNN/README.md) | [PyTorch](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/5-NLP/16-RNN/RNNPyTorch.ipynb) / [TensorFlow](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/5-NLP/16-RNN/RNNTF.ipynb) | |
|
||||
| 17 | [生成型循環網路](./lessons/5-NLP/17-GenerativeNetworks/README.md) | [PyTorch](https://microsoft.github.io/AI-For-Beginners/lessons/5-NLP/17-GenerativeNetworks/GenerativePyTorch.md) / [TensorFlow](https://microsoft.github.io/AI-For-Beginners/lessons/5-NLP/17-GenerativeNetworks/GenerativeTF.md) | [實驗室](./lessons/5-NLP/17-GenerativeNetworks/lab/README.md) |
|
||||
| 18 | [Transformer. BERT.](./lessons/5-NLP/18-Transformers/READMEtransformers.md) | [PyTorch](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/5-NLP/18-Transformers/TransformersPyTorch.ipynb) /[TensorFlow](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/5-NLP/18-Transformers/TransformersTF.ipynb) | |
|
||||
| 19 | [命名實體識別 (NER)](./lessons/5-NLP/19-NER/README.md) | [TensorFlow](https://microsoft.github.io/AI-For-Beginners/lessons/5-NLP/19-NER/NER-TF.ipynb) | [實驗室](./lessons/5-NLP/19-NER/lab/README.md) |
|
||||
| 20 | [大型語言模型、提示編程與少樣本任務](./lessons/5-NLP/20-LangModels/READMELargeLang.md) | [PyTorch](https://microsoft.github.io/AI-For-Beginners/lessons/5-NLP/20-LangModels/GPT-PyTorch.ipynb) | |
|
||||
我們不會在這個課程中涵蓋：
|
||||
|
||||
> [在 Microsoft Learn 集合中找到本課程的所有額外資源](https://learn.microsoft.com/en-us/collections/7w28iy2xrqzdj0?WT.mc_id=academic-77998-bethanycheum)
|
||||
|
||||
* **AI 在商業中的應用案例**。建議參加 [商業用戶的 AI 入門](https://docs.microsoft.com/learn/paths/introduction-ai-for-business-users/?WT.mc_id=academic-77998-bethanycheum) 學習路徑，或 [AI 商業學院](https://www.microsoft.com/ai/ai-business-school/?WT.mc_id=academic-77998-bethanycheum)，由 [INSEAD](https://www.insead.edu/) 合作開發。
|
||||
* **經典機器學習**，這在我們的 [機器學習入門課程](http://github.com/Microsoft/ML-for-Beginners) 中有詳細描述。
|
||||
* 使用 **[認知服務](https://azure.microsoft.com/services/cognitive-services/?WT.mc_id=academic-77998-bethanycheum)** 構建的實際 AI 應用。建議從 Microsoft Learn 的模組開始，例如 [視覺](https://docs.microsoft.com/learn/paths/create-computer-vision-solutions-azure-cognitive-services/?WT.mc_id=academic-77998-bethanycheum)、[自然語言處理](https://docs.microsoft.com/learn/paths/explore-natural-language-processing/?WT.mc_id=academic-77998-bethanycheum)、**[使用 Azure OpenAI 服務的生成式 AI](https://learn.microsoft.com/en-us/training/paths/develop-ai-solutions-azure-openai/?WT.mc_id=academic-77998-bethanycheum)** 等。
|
||||
* 特定的 ML **雲框架**，例如 [Azure Machine Learning](https://azure.microsoft.com/services/machine-learning/?WT.mc_id=academic-77998-bethanycheum)、[Microsoft Fabric](https://learn.microsoft.com/en-us/training/paths/get-started-fabric/?WT.mc_id=academic-77998-bethanycheum) 或 [Azure Databricks](https://docs.microsoft.com/learn/paths/data-engineer-azure-databricks?WT.mc_id=academic-77998-bethanycheum)。建議參考 [使用 Azure Machine Learning 構建和運營機器學習解決方案](https://docs.microsoft.com/learn/paths/build-ai-solutions-with-azure-ml-service/?WT.mc_id=academic-77998-bethanycheum) 和 [使用 Azure Databricks 構建和運營機器學習解決方案](https://docs.microsoft.com/learn/paths/build-operate-machine-learning-solutions-azure-databricks/?WT.mc_id=academic-77998-bethanycheum) 學習路徑。
|
||||
* **對話式 AI** 和 **聊天機器人**。有一個單獨的 [創建對話式 AI 解決方案](https://docs.microsoft.com/learn/paths/create-conversational-ai-solutions/?WT.mc_id=academic-77998-bethanycheum) 學習路徑，您也可以參考 [這篇部落格文章](https://soshnikov.com/azure/hello-bot-conversational-ai-on-microsoft-platform/) 獲取更多細節。
|
||||
* **深度學習的數學基礎**。建議參考 Ian Goodfellow、Yoshua Bengio 和 Aaron Courville 所著的 [Deep Learning](https://www.amazon.com/Deep-Learning-Adaptive-Computation-Machine/dp/0262035618)，該書也可在線上獲取：[https://www.deeplearningbook.org/](https://www.deeplearningbook.org/)。
|
||||
|
||||
如果您想要輕鬆入門 _雲端 AI_ 主題，可以考慮參加 [在 Azure 上開始人工智慧之旅](https://docs.microsoft.com/learn/paths/get-started-with-artificial-intelligence-on-azure/?WT.mc_id=academic-77998-bethanycheum) 學習路徑。
|
||||
|
||||
# 課程內容
|
||||
|
||||
| | 課程連結 | PyTorch/Keras/TensorFlow | 實驗室 |
|
||||
| :-: | :------------------------------------------------------------------------------------------------------------------------------------------: | :---------------------------------------------------------------------------------------------: | ------------------------------------------------------------------------------ |
|
||||
| 0 | [課程設置](./lessons/0-course-setup/setup.md) | [設置您的開發環境](./lessons/0-course-setup/how-to-run.md) | |
|
||||
| I | [**AI 簡介**](./lessons/1-Intro/README.md) | | |
|
||||
| 01 | [AI 的簡介與歷史](./lessons/1-Intro/README.md) | - | - |
|
||||
| II | **符號 AI** |
|
||||
| 02 | [知識表徵與專家系統](./lessons/2-Symbolic/README.md) | [專家系統](./lessons/2-Symbolic/Animals.ipynb) / [本體論](./lessons/2-Symbolic/FamilyOntology.ipynb) /[概念圖](./lessons/2-Symbolic/MSConceptGraph.ipynb) | |
|
||||
| III | [**神經網路簡介**](./lessons/3-NeuralNetworks/README.md) |||
|
||||
| 03 | [感知器](./lessons/3-NeuralNetworks/03-Perceptron/README.md) | [筆記本](./lessons/3-NeuralNetworks/03-Perceptron/Perceptron.ipynb) | [實驗室](./lessons/3-NeuralNetworks/03-Perceptron/lab/README.md) |
|
||||
| 04 | [多層感知器與創建我們自己的框架](./lessons/3-NeuralNetworks/04-OwnFramework/README.md) | [筆記本](./lessons/3-NeuralNetworks/04-OwnFramework/OwnFramework.ipynb) | [實驗室](./lessons/3-NeuralNetworks/04-OwnFramework/lab/README.md) |
|
||||
| 05 | [框架介紹（PyTorch/TensorFlow）與過擬合](./lessons/3-NeuralNetworks/05-Frameworks/README.md) | [PyTorch](./lessons/3-NeuralNetworks/05-Frameworks/IntroPyTorch.ipynb) / [Keras](./lessons/3-NeuralNetworks/05-Frameworks/IntroKeras.ipynb) / [TensorFlow](./lessons/3-NeuralNetworks/05-Frameworks/IntroKerasTF.ipynb) | [實驗室](./lessons/3-NeuralNetworks/05-Frameworks/lab/README.md) |
|
||||
| IV | [**電腦視覺**](./lessons/4-ComputerVision/README.md) | [PyTorch](https://docs.microsoft.com/learn/modules/intro-computer-vision-pytorch/?WT.mc_id=academic-77998-cacaste) / [TensorFlow](https://docs.microsoft.com/learn/modules/intro-computer-vision-TensorFlow/?WT.mc_id=academic-77998-cacaste)| [探索 Microsoft Azure 上的電腦視覺](https://learn.microsoft.com/en-us/collections/7w28iy2xrqzdj0?WT.mc_id=academic-77998-bethanycheum) |
|
||||
| 06 | [電腦視覺入門. OpenCV](./lessons/4-ComputerVision/06-IntroCV/README.md) | [筆記本](./lessons/4-ComputerVision/06-IntroCV/OpenCV.ipynb) | [實驗室](./lessons/4-ComputerVision/06-IntroCV/lab/README.md) |
|
||||
| 07 | [卷積神經網絡](./lessons/4-ComputerVision/07-ConvNets/README.md) & [CNN 架構](./lessons/4-ComputerVision/07-ConvNets/CNN_Architectures.md) | [PyTorch](./lessons/4-ComputerVision/07-ConvNets/ConvNetsPyTorch.ipynb) /[TensorFlow](./lessons/4-ComputerVision/07-ConvNets/ConvNetsTF.ipynb) | [實驗室](./lessons/4-ComputerVision/07-ConvNets/lab/README.md) |
|
||||
| 08 | [預訓練網絡與遷移學習](./lessons/4-ComputerVision/08-TransferLearning/README.md) 和 [訓練技巧](./lessons/4-ComputerVision/08-TransferLearning/TrainingTricks.md) | [PyTorch](./lessons/4-ComputerVision/08-TransferLearning/TransferLearningPyTorch.ipynb) / [TensorFlow](./lessons/3-NeuralNetworks/05-Frameworks/IntroKerasTF.ipynb) | [實驗室](./lessons/4-ComputerVision/08-TransferLearning/lab/README.md) |
|
||||
| 09 | [自編碼器與變分自編碼器 (VAEs)](./lessons/4-ComputerVision/09-Autoencoders/README.md) | [PyTorch](./lessons/4-ComputerVision/09-Autoencoders/AutoEncodersPyTorch.ipynb) / [TensorFlow](./lessons/4-ComputerVision/09-Autoencoders/AutoencodersTF.ipynb) | |
|
||||
| 10 | [生成對抗網絡 (GANs) 與藝術風格遷移](./lessons/4-ComputerVision/10-GANs/README.md) | [PyTorch](./lessons/4-ComputerVision/10-GANs/GANPyTorch.ipynb) / [TensorFlow](./lessons/4-ComputerVision/10-GANs/GANTF.ipynb) | |
|
||||
| 11 | [物體檢測](./lessons/4-ComputerVision/11-ObjectDetection/README.md) | [TensorFlow](./lessons/4-ComputerVision/11-ObjectDetection/ObjectDetection.ipynb) | [實驗室](./lessons/4-ComputerVision/11-ObjectDetection/lab/README.md) |
|
||||
| 12 | [語義分割. U-Net](./lessons/4-ComputerVision/12-Segmentation/README.md) | [PyTorch](./lessons/4-ComputerVision/12-Segmentation/SemanticSegmentationPytorch.ipynb) / [TensorFlow](./lessons/4-ComputerVision/12-Segmentation/SemanticSegmentationTF.ipynb) | |
|
||||
| V | [**自然語言處理**](./lessons/5-NLP/README.md) | [PyTorch](https://docs.microsoft.com/learn/modules/intro-natural-language-processing-pytorch/?WT.mc_id=academic-77998-cacaste) /[TensorFlow](https://docs.microsoft.com/learn/modules/intro-natural-language-processing-TensorFlow/?WT.mc_id=academic-77998-cacaste) | [探索 Microsoft Azure 上的自然語言處理](https://learn.microsoft.com/en-us/collections/7w28iy2xrqzdj0?WT.mc_id=academic-77998-bethanycheum)|
|
||||
| 13 | [文本表示. Bow/TF-IDF](./lessons/5-NLP/13-TextRep/README.md) | [PyTorch](./lessons/5-NLP/13-TextRep/TextRepresentationPyTorch.ipynb) / [TensorFlow](./lessons/5-NLP/13-TextRep/TextRepresentationTF.ipynb) | |
|
||||
| 14 | [語義詞嵌入. Word2Vec 和 GloVe](./lessons/5-NLP/14-Embeddings/README.md) | [PyTorch](./lessons/5-NLP/14-Embeddings/EmbeddingsPyTorch.ipynb) / [TensorFlow](./lessons/5-NLP/14-Embeddings/EmbeddingsTF.ipynb) | |
|
||||
| 15 | [語言建模. 訓練自己的嵌入](./lessons/5-NLP/15-LanguageModeling/README.md) | [PyTorch](./lessons/5-NLP/15-LanguageModeling/CBoW-PyTorch.ipynb) / [TensorFlow](./lessons/5-NLP/15-LanguageModeling/CBoW-TF.ipynb) | [實驗室](./lessons/5-NLP/15-LanguageModeling/lab/README.md) |
|
||||
| 16 | [循環神經網絡 (RNN)](./lessons/5-NLP/16-RNN/README.md) | [PyTorch](./lessons/5-NLP/16-RNN/RNNPyTorch.ipynb) / [TensorFlow](./lessons/5-NLP/16-RNN/RNNTF.ipynb) | |
|
||||
| 17 | [生成式循環網絡](./lessons/5-NLP/17-GenerativeNetworks/README.md) | [PyTorch](./lessons/5-NLP/17-GenerativeNetworks/GenerativePyTorch.md) / [TensorFlow](./lessons/5-NLP/17-GenerativeNetworks/GenerativeTF.md) | [實驗室](./lessons/5-NLP/17-GenerativeNetworks/lab/README.md) |
|
||||
| 18 | [Transformer. BERT.](./lessons/5-NLP/18-Transformers/READMEtransformers.md) | [PyTorch](./lessons/5-NLP/18-Transformers/TransformersPyTorch.ipynb) /[TensorFlow](./lessons/5-NLP/18-Transformers/TransformersTF.ipynb) | |
|
||||
| 19 | [命名實體識別 (NER)](./lessons/5-NLP/19-NER/README.md) | [TensorFlow](./lessons/5-NLP/19-NER/NER-TF.ipynb) | [實驗室](./lessons/5-NLP/19-NER/lab/README.md) |
|
||||
| 20 | [大型語言模型、提示編程與少樣本任務](./lessons/5-NLP/20-LangModels/READMELargeLang.md) | [PyTorch](./lessons/5-NLP/20-LangModels/GPT-PyTorch.ipynb) | |
|
||||
| VI | **其他 AI 技術** || |
|
||||
| 21 | [遺傳演算法](./lessons/6-Other/21-GeneticAlgorithms/README.md) | [Notebook](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/6-Other/21-GeneticAlgorithms/Genetic.ipynb) | |
|
||||
| 22 | [深度強化學習](./lessons/6-Other/22-DeepRL/README.md) | [PyTorch](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/6-Other/22-DeepRL/CartPole-RL-PyTorch.ipynb) /[TensorFlow](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/6-Other/22-DeepRL/CartPole-RL-TF.ipynb) | [實驗室](./lessons/6-Other/22-DeepRL/lab/README.md) |
|
||||
| 21 | [遺傳算法](./lessons/6-Other/21-GeneticAlgorithms/README.md) | [筆記本](./lessons/6-Other/21-GeneticAlgorithms/Genetic.ipynb) | |
|
||||
| 22 | [深度強化學習](./lessons/6-Other/22-DeepRL/README.md) | [PyTorch](./lessons/6-Other/22-DeepRL/CartPole-RL-PyTorch.ipynb) /[TensorFlow](./lessons/6-Other/22-DeepRL/CartPole-RL-TF.ipynb) | [實驗室](./lessons/6-Other/22-DeepRL/lab/README.md) |
|
||||
| 23 | [多代理系統](./lessons/6-Other/23-MultiagentSystems/README.md) | | |
|
||||
| VII | **AI 倫理** | | |
|
||||
| 24 | [AI 倫理與負責任的 AI](./lessons/7-Ethics/README.md) | [Microsoft Learn: 負責任的 AI 原則](https://docs.microsoft.com/learn/paths/responsible-ai-business-principles/?WT.mc_id=academic-77998-cacaste) | |
|
||||
| IX | **額外內容** | | |
|
||||
| 25 | [多模態網路、CLIP 和 VQGAN](./lessons/X-Extras/X1-MultiModal/README.md) | [Notebook](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/X-Extras/X1-MultiModal/Clip.ipynb) | |
|
||||
| IX | **附加內容** | | |
|
||||
| 25 | [多模態網絡, CLIP 和 VQGAN](./lessons/X-Extras/X1-MultiModal/README.md) | [筆記本](./lessons/X-Extras/X1-MultiModal/Clip.ipynb) | |
|
||||
|
||||
## 每個課程包含
|
||||
## 每節課包含
|
||||
|
||||
* 預讀材料
|
||||
* 可執行的 Jupyter Notebook，通常針對特定框架（**PyTorch** 或 **TensorFlow**）。可執行的 Notebook 也包含大量理論內容，因此為了理解主題，您需要至少完成一個版本的 Notebook（PyTorch 或 TensorFlow）。
|
||||
* **實驗室**：部分主題提供實驗室，讓您有機會將所學應用於特定問題。
|
||||
* 部分章節包含指向 [**MS Learn**](https://learn.microsoft.com/en-us/collections/7w28iy2xrqzdj0?WT.mc_id=academic-77998-bethanycheum) 模組的連結，涵蓋相關主題。
|
||||
* 預讀材料
|
||||
* 可執行的 Jupyter 筆記本，通常針對特定框架（**PyTorch** 或 **TensorFlow**）。可執行的筆記本還包含大量理論材料，因此為了理解主題，您需要至少閱讀一個版本的筆記本（PyTorch 或 TensorFlow）。
|
||||
* **實驗室**（部分主題提供），讓您有機會將所學應用於具體問題。
|
||||
* 某些部分包含指向 [**MS Learn**](https://learn.microsoft.com/en-us/collections/7w28iy2xrqzdj0?WT.mc_id=academic-77998-bethanycheum) 模組的鏈接，涵蓋相關主題。
|
||||
|
||||
## 開始學習
|
||||
|
||||
- 我們已經創建了一個 [設置課程](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/0-course-setup/setup.md)，幫助您設置開發環境。
|
||||
- 對於教育者，我們也創建了一個 [課程設置指南](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/0-course-setup/for-teachers.md)！
|
||||
- 如何 [在 VSCode 或 Codepace 中運行代碼](https://github.com/microsoft/AI-For-Beginners/blob/main/lessons/0-course-setup/how-to-run.md)
|
||||
- 我們已經創建了一個 [設置課程](./lessons/0-course-setup/setup.md) 幫助您設置開發環境。 - 對於教育者，我們還創建了一個 [課程設置指南](./lessons/0-course-setup/for-teachers.md)！
|
||||
- 如何 [在 VSCode 或 Codepace 中運行代碼](./lessons/0-course-setup/how-to-run.md)
|
||||
|
||||
請按照以下步驟操作：
|
||||
|
||||
1. Fork 此倉庫：點擊頁面右上角的 "Fork" 按鈕。
|
||||
2. 克隆倉庫：`git clone https://github.com/microsoft/AI-For-Beginners.git`
|
||||
3. 別忘了給這個倉庫加星標 (🌟)，以便日後更容易找到。
|
||||
分叉此倉庫：點擊此頁面右上角的 "Fork" 按鈕。
|
||||
|
||||
克隆倉庫：`git clone https://github.com/microsoft/AI-For-Beginners.git`
|
||||
|
||||
別忘了給這個倉庫加星標 (🌟)，以便以後更容易找到。
|
||||
|
||||
## 與其他學習者交流
|
||||
|
||||
加入我們的 [官方 AI Discord 伺服器](https://aka.ms/genai-discord?WT.mc_id=academic-105485-bethanycheum)，與其他學習者交流並獲得支持。
|
||||
|
||||
如果您在構建過程中有產品反饋或問題，請訪問我們的 [Azure AI Foundry 開發者論壇](https://aka.ms/foundry/forum)。
|
||||
如果您在構建過程中有產品反饋或問題，請訪問我們的 [Azure AI Foundry 開發者論壇](https://aka.ms/foundry/forum)
|
||||
|
||||
## 測驗
|
||||
> **關於測驗的注意事項**：所有測驗都存放在 etc\quiz-app 的 Quiz-app 資料夾中，並從課程中進行連結。測驗應用程式可以在本地執行或部署到 Azure；請按照 `quiz-app` 資料夾中的指示操作。測驗正在逐步進行本地化。
|
||||
## 徵求協助
|
||||
> **關於測驗的一點說明**：所有測驗都存放在 Quiz-app 資料夾中，路徑為 etc\quiz-app，或者可以[在線訪問這裡](https://ff-quizzes.netlify.app/)。這些測驗已經在課程中進行了連結。測驗應用程式可以在本地運行，也可以部署到 Azure；請按照 `quiz-app` 資料夾中的指示操作。測驗內容正在逐步進行本地化。
|
||||
## 尋求協助
|
||||
|
||||
您有建議或發現拼寫或程式碼錯誤嗎？請提出問題或建立拉取請求。
|
||||
您有任何建議或發現拼寫或程式碼錯誤嗎？請提出問題或建立拉取請求。
|
||||
|
||||
## 特別感謝
|
||||
|
||||
|
|
@ -150,5 +162,7 @@ CO_OP_TRANSLATOR_METADATA:
|
|||
- [精通 GitHub Copilot 用於 C#/.NET 開發者](https://github.com/microsoft/mastering-github-copilot-for-dotnet-csharp-developers)
|
||||
- [選擇你的 Copilot 冒險](https://github.com/microsoft/CopilotAdventures)
|
||||
|
||||
---
|
||||
|
||||
**免責聲明**：
|
||||
本文件使用 AI 翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。應以原文文件作為權威來源。對於關鍵資訊，建議尋求專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或錯誤解釋概不負責。
|
||||
本文件已使用 AI 翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。儘管我們致力於提供準確的翻譯，請注意自動翻譯可能包含錯誤或不準確之處。原始文件的母語版本應被視為權威來源。對於關鍵資訊，建議使用專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或錯誤解釋不承擔責任。
|
||||
|
|
@ -0,0 +1,478 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"collapsed": true
|
||||
},
|
||||
"source": [
|
||||
"# 實現動物專家系統\n",
|
||||
"\n",
|
||||
"範例來源：[AI for Beginners Curriculum](http://github.com/microsoft/ai-for-beginners)。\n",
|
||||
"\n",
|
||||
"在此範例中，我們將實現一個簡單的基於知識的系統，根據一些外部特徵來判斷動物。該系統可以用以下的 AND-OR 樹來表示（這只是整個樹的一部分，我們可以輕鬆地添加更多規則）：\n",
|
||||
"\n",
|
||||
"\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 我們自己的專家系統殼層與反向推理\n",
|
||||
"\n",
|
||||
"讓我們嘗試基於生成規則定義一種簡單的知識表示語言。我們將使用 Python 類作為關鍵字來定義規則。基本上會有三種類型的類別：\n",
|
||||
"* `Ask` 代表需要向使用者提問的問題。它包含可能的答案集合。\n",
|
||||
"* `If` 代表一條規則，它只是用來存儲規則內容的語法糖。\n",
|
||||
"* `AND`/`OR` 是用來表示樹的 AND/OR 分支的類別。它們僅存儲內部的參數列表。為了簡化程式碼，所有功能都定義在父類別 `Content` 中。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 1,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"class Ask():\n",
|
||||
" def __init__(self,choices=['y','n']):\n",
|
||||
" self.choices = choices\n",
|
||||
" def ask(self):\n",
|
||||
" if max([len(x) for x in self.choices])>1:\n",
|
||||
" for i,x in enumerate(self.choices):\n",
|
||||
" print(\"{0}. {1}\".format(i,x),flush=True)\n",
|
||||
" x = int(input())\n",
|
||||
" return self.choices[x]\n",
|
||||
" else:\n",
|
||||
" print(\"/\".join(self.choices),flush=True)\n",
|
||||
" return input()\n",
|
||||
"\n",
|
||||
"class Content():\n",
|
||||
" def __init__(self,x):\n",
|
||||
" self.x=x\n",
|
||||
" \n",
|
||||
"class If(Content):\n",
|
||||
" pass\n",
|
||||
"\n",
|
||||
"class AND(Content):\n",
|
||||
" pass\n",
|
||||
"\n",
|
||||
"class OR(Content):\n",
|
||||
" pass"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"在我們的系統中，工作記憶將包含作為**屬性-值對**的**事實**列表。知識庫可以定義為一個大的字典，將行動（應插入工作記憶中的新事實）映射到條件，這些條件以 AND-OR 表達式表示。此外，一些事實可以被`詢問`。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 2,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"rules = {\n",
|
||||
" 'default': Ask(['y','n']),\n",
|
||||
" 'color' : Ask(['red-brown','black and white','other']),\n",
|
||||
" 'pattern' : Ask(['dark stripes','dark spots']),\n",
|
||||
" 'mammal': If(OR(['hair','gives milk'])),\n",
|
||||
" 'carnivor': If(OR([AND(['sharp teeth','claws','forward-looking eyes']),'eats meat'])),\n",
|
||||
" 'ungulate': If(['mammal',OR(['has hooves','chews cud'])]),\n",
|
||||
" 'bird': If(OR(['feathers',AND(['flies','lies eggs'])])),\n",
|
||||
" 'animal:monkey' : If(['mammal','carnivor','color:red-brown','pattern:dark spots']),\n",
|
||||
" 'animal:tiger' : If(['mammal','carnivor','color:red-brown','pattern:dark stripes']),\n",
|
||||
" 'animal:giraffe' : If(['ungulate','long neck','long legs','pattern:dark spots']),\n",
|
||||
" 'animal:zebra' : If(['ungulate','pattern:dark stripes']),\n",
|
||||
" 'animal:ostrich' : If(['bird','long nech','color:black and white','cannot fly']),\n",
|
||||
" 'animal:pinguin' : If(['bird','swims','color:black and white','cannot fly']),\n",
|
||||
" 'animal:albatross' : If(['bird','flies well'])\n",
|
||||
"}"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"為了執行反向推理，我們將定義 `Knowledgebase` 類別。它將包含：\n",
|
||||
"* 工作中的 `memory` - 一個將屬性映射到值的字典\n",
|
||||
"* 知識庫中的 `rules` - 以上述定義的格式表示\n",
|
||||
"\n",
|
||||
"兩個主要的方法是：\n",
|
||||
"* `get` 用於獲取屬性的值，必要時執行推理。例如，`get('color')` 將獲取顏色槽的值（如果需要，會詢問並將值存儲在工作記憶中以供後續使用）。如果我們詢問 `get('color:blue')`，它將詢問顏色，然後根據顏色返回 `y`/`n` 值。\n",
|
||||
"* `eval` 執行實際的推理，即遍歷 AND/OR 樹，評估子目標等。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"class KnowledgeBase():\n",
|
||||
" def __init__(self,rules):\n",
|
||||
" self.rules = rules\n",
|
||||
" self.memory = {}\n",
|
||||
" \n",
|
||||
" def get(self,name):\n",
|
||||
" if ':' in name:\n",
|
||||
" k,v = name.split(':')\n",
|
||||
" vv = self.get(k)\n",
|
||||
" return 'y' if v==vv else 'n'\n",
|
||||
" if name in self.memory.keys():\n",
|
||||
" return self.memory[name]\n",
|
||||
" for fld in self.rules.keys():\n",
|
||||
" if fld==name or fld.startswith(name+\":\"):\n",
|
||||
" # print(\" + proving {}\".format(fld))\n",
|
||||
" value = 'y' if fld==name else fld.split(':')[1]\n",
|
||||
" res = self.eval(self.rules[fld],field=name)\n",
|
||||
" if res!='y' and res!='n' and value=='y':\n",
|
||||
" self.memory[name] = res\n",
|
||||
" return res\n",
|
||||
" if res=='y':\n",
|
||||
" self.memory[name] = value\n",
|
||||
" return value\n",
|
||||
" # field is not found, using default\n",
|
||||
" res = self.eval(self.rules['default'],field=name)\n",
|
||||
" self.memory[name]=res\n",
|
||||
" return res\n",
|
||||
" \n",
|
||||
" def eval(self,expr,field=None):\n",
|
||||
" # print(\" + eval {}\".format(expr))\n",
|
||||
" if isinstance(expr,Ask):\n",
|
||||
" print(field)\n",
|
||||
" return expr.ask()\n",
|
||||
" elif isinstance(expr,If):\n",
|
||||
" return self.eval(expr.x)\n",
|
||||
" elif isinstance(expr,AND) or isinstance(expr,list):\n",
|
||||
" expr = expr.x if isinstance(expr,AND) else expr\n",
|
||||
" for x in expr:\n",
|
||||
" if self.eval(x)=='n':\n",
|
||||
" return 'n'\n",
|
||||
" return 'y'\n",
|
||||
" elif isinstance(expr,OR):\n",
|
||||
" for x in expr.x:\n",
|
||||
" if self.eval(x)=='y':\n",
|
||||
" return 'y'\n",
|
||||
" return 'n'\n",
|
||||
" elif isinstance(expr,str):\n",
|
||||
" return self.get(expr)\n",
|
||||
" else:\n",
|
||||
" print(\"Unknown expr: {}\".format(expr))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在讓我們定義我們的動物知識庫並進行諮詢。請注意，此操作將向您提問。您可以通過輸入 `y`/`n` 來回答是非問題，或者通過指定數字（0..N）來回答具有較多選擇的問題。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"hair\n",
|
||||
"y/n\n",
|
||||
"sharp teeth\n",
|
||||
"y/n\n",
|
||||
"claws\n",
|
||||
"y/n\n",
|
||||
"forward-looking eyes\n",
|
||||
"y/n\n",
|
||||
"color\n",
|
||||
"0. red-brown\n",
|
||||
"1. black and white\n",
|
||||
"2. other\n",
|
||||
"has hooves\n",
|
||||
"y/n\n",
|
||||
"long neck\n",
|
||||
"y/n\n",
|
||||
"long legs\n",
|
||||
"y/n\n",
|
||||
"pattern\n",
|
||||
"0. dark stripes\n",
|
||||
"1. dark spots\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"'giraffe'"
|
||||
]
|
||||
},
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"kb = KnowledgeBase(rules)\n",
|
||||
"kb.get('animal')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 使用 PyKnow 進行前向推理\n",
|
||||
"\n",
|
||||
"在接下來的範例中，我們將嘗試使用一個知識表示的函式庫 [PyKnow](https://github.com/buguroo/pyknow/) 來實現前向推理。**PyKnow** 是一個用於在 Python 中建立前向推理系統的函式庫，其設計類似於經典的舊系統 [CLIPS](http://www.clipsrules.net/index.html)。\n",
|
||||
"\n",
|
||||
"我們也可以自己實現前向鏈結，這並不會有太大的困難，但簡單的實現通常效率不高。為了更有效地進行規則匹配，會使用一種特殊的演算法 [Rete](https://en.wikipedia.org/wiki/Rete_algorithm)。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Collecting git+https://github.com/buguroo/pyknow/\n",
|
||||
" Cloning https://github.com/buguroo/pyknow/ to /tmp/pip-req-build-3cqeulyl\n",
|
||||
" Running command git clone --filter=blob:none --quiet https://github.com/buguroo/pyknow/ /tmp/pip-req-build-3cqeulyl\n",
|
||||
" Resolved https://github.com/buguroo/pyknow/ to commit 48818336f2e9a126f1964f2d8dc22d37ff800fe8\n",
|
||||
" Preparing metadata (setup.py) ... \u001b[?25ldone\n",
|
||||
"\u001b[?25hCollecting frozendict==1.2\n",
|
||||
" Using cached frozendict-1.2.tar.gz (2.6 kB)\n",
|
||||
" Preparing metadata (setup.py) ... \u001b[?25ldone\n",
|
||||
"\u001b[?25hCollecting schema==0.6.7\n",
|
||||
" Using cached schema-0.6.7-py2.py3-none-any.whl (14 kB)\n",
|
||||
"Building wheels for collected packages: pyknow, frozendict\n",
|
||||
" Building wheel for pyknow (setup.py) ... \u001b[?25ldone\n",
|
||||
"\u001b[?25h Created wheel for pyknow: filename=pyknow-1.7.0-py3-none-any.whl size=34228 sha256=b7de5b09292c4007667c72f69b98d5a1b5f7324ff15f9dd8e077c3d5f7aade42\n",
|
||||
" Stored in directory: /tmp/pip-ephem-wheel-cache-k7jpave7/wheels/81/1a/d3/f6c15dbe1955598a37755215f2a10449e7418500d7bd4b9508\n",
|
||||
" Building wheel for frozendict (setup.py) ... \u001b[?25ldone\n",
|
||||
"\u001b[?25h Created wheel for frozendict: filename=frozendict-1.2-py3-none-any.whl size=3148 sha256=2863d55c240d2409cddf05ccfe600591f8478681549fc97555c47c90dc6bb160\n",
|
||||
" Stored in directory: /home/rg/.cache/pip/wheels/49/ac/f8/cb8120244e710bdb479c86198b03c7b08c3c2d3d2bf448fd6e\n",
|
||||
"Successfully built pyknow frozendict\n",
|
||||
"Installing collected packages: schema, frozendict, pyknow\n",
|
||||
"Successfully installed frozendict-1.2 pyknow-1.7.0 schema-0.6.7\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"import sys\n",
|
||||
"!{sys.executable} -m pip install git+https://github.com/buguroo/pyknow/"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 13,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"from pyknow import *\n",
|
||||
"#import pyknow"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們將把系統定義為一個繼承 `KnowledgeEngine` 的類別。每條規則由一個帶有 `@Rule` 註解的單獨函數定義，該註解指定了規則應該觸發的時機。在規則內部，我們可以使用 `declare` 函數添加新的事實，添加這些事實將導致前向推理引擎調用更多的規則。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 14,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"class Animals(KnowledgeEngine):\n",
|
||||
" @Rule(OR(\n",
|
||||
" AND(Fact('sharp teeth'),Fact('claws'),Fact('forward looking eyes')),\n",
|
||||
" Fact('eats meat')))\n",
|
||||
" def cornivor(self):\n",
|
||||
" self.declare(Fact('carnivor'))\n",
|
||||
" \n",
|
||||
" @Rule(OR(Fact('hair'),Fact('gives milk')))\n",
|
||||
" def mammal(self):\n",
|
||||
" self.declare(Fact('mammal'))\n",
|
||||
"\n",
|
||||
" @Rule(Fact('mammal'),\n",
|
||||
" OR(Fact('has hooves'),Fact('chews cud')))\n",
|
||||
" def hooves(self):\n",
|
||||
" self.declare('ungulate')\n",
|
||||
" \n",
|
||||
" @Rule(OR(Fact('feathers'),AND(Fact('flies'),Fact('lays eggs'))))\n",
|
||||
" def bird(self):\n",
|
||||
" self.declare('bird')\n",
|
||||
" \n",
|
||||
" @Rule(Fact('mammal'),Fact('carnivor'),\n",
|
||||
" Fact(color='red-brown'),\n",
|
||||
" Fact(pattern='dark spots'))\n",
|
||||
" def monkey(self):\n",
|
||||
" self.declare(Fact(animal='monkey'))\n",
|
||||
"\n",
|
||||
" @Rule(Fact('mammal'),Fact('carnivor'),\n",
|
||||
" Fact(color='red-brown'),\n",
|
||||
" Fact(pattern='dark stripes'))\n",
|
||||
" def tiger(self):\n",
|
||||
" self.declare(Fact(animal='tiger'))\n",
|
||||
"\n",
|
||||
" @Rule(Fact('ungulate'),\n",
|
||||
" Fact('long neck'),\n",
|
||||
" Fact('long legs'),\n",
|
||||
" Fact(pattern='dark spots'))\n",
|
||||
" def giraffe(self):\n",
|
||||
" self.declare(Fact(animal='giraffe'))\n",
|
||||
"\n",
|
||||
" @Rule(Fact('ungulate'),\n",
|
||||
" Fact(pattern='dark stripes'))\n",
|
||||
" def zebra(self):\n",
|
||||
" self.declare(Fact(animal='zebra'))\n",
|
||||
"\n",
|
||||
" @Rule(Fact('bird'),\n",
|
||||
" Fact('long neck'),\n",
|
||||
" Fact('cannot fly'),\n",
|
||||
" Fact(color='black and white'))\n",
|
||||
" def straus(self):\n",
|
||||
" self.declare(Fact(animal='ostrich'))\n",
|
||||
"\n",
|
||||
" @Rule(Fact('bird'),\n",
|
||||
" Fact('swims'),\n",
|
||||
" Fact('cannot fly'),\n",
|
||||
" Fact(color='black and white'))\n",
|
||||
" def pinguin(self):\n",
|
||||
" self.declare(Fact(animal='pinguin'))\n",
|
||||
"\n",
|
||||
" @Rule(Fact('bird'),\n",
|
||||
" Fact('flies well'))\n",
|
||||
" def albatros(self):\n",
|
||||
" self.declare(Fact(animal='albatross'))\n",
|
||||
" \n",
|
||||
" @Rule(Fact(animal=MATCH.a))\n",
|
||||
" def print_result(self,a):\n",
|
||||
" print('Animal is {}'.format(a))\n",
|
||||
" \n",
|
||||
" def factz(self,l):\n",
|
||||
" for x in l:\n",
|
||||
" self.declare(x)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"一旦我們定義了一個知識庫，我們會用一些初始事實填充工作記憶，然後調用 `run()` 方法來執行推理。結果你可以看到新的推導事實被添加到工作記憶中，包括關於動物的最終事實（如果我們正確設置了所有初始事實）。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 15,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Animal is tiger\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"FactList([(0, InitialFact()),\n",
|
||||
" (1, Fact(color='red-brown')),\n",
|
||||
" (2, Fact(pattern='dark stripes')),\n",
|
||||
" (3, Fact('sharp teeth')),\n",
|
||||
" (4, Fact('claws')),\n",
|
||||
" (5, Fact('forward looking eyes')),\n",
|
||||
" (6, Fact('gives milk')),\n",
|
||||
" (7, Fact('mammal')),\n",
|
||||
" (8, Fact('carnivor')),\n",
|
||||
" (9, Fact(animal='tiger'))])"
|
||||
]
|
||||
},
|
||||
"execution_count": 15,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"ex1 = Animals()\n",
|
||||
"ex1.reset()\n",
|
||||
"ex1.factz([\n",
|
||||
" Fact(color='red-brown'),\n",
|
||||
" Fact(pattern='dark stripes'),\n",
|
||||
" Fact('sharp teeth'),\n",
|
||||
" Fact('claws'),\n",
|
||||
" Fact('forward looking eyes'),\n",
|
||||
" Fact('gives milk')])\n",
|
||||
"ex1.run()\n",
|
||||
"ex1.facts"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": []
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件已使用 AI 翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。儘管我們努力確保翻譯的準確性，但請注意，自動翻譯可能包含錯誤或不準確之處。原始文件的母語版本應被視為權威來源。對於關鍵信息，建議使用專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或錯誤解釋不承擔責任。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"kernelspec": {
|
||||
"display_name": "Python 3.7.4 64-bit (conda)",
|
||||
"metadata": {
|
||||
"interpreter": {
|
||||
"hash": "86193a1ab0ba47eac1c69c1756090baa3b420b3eea7d4aafab8b85f8b312f0c5"
|
||||
}
|
||||
},
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.11.2"
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "ab2bd97b0453415b89a469284609a8ce",
|
||||
"translation_date": "2025-08-31T10:07:14+00:00",
|
||||
"source_file": "lessons/2-Symbolic/Animals.ipynb",
|
||||
"language_code": "tw"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 2
|
||||
}
|
||||
|
|
@ -0,0 +1,595 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"collapsed": true
|
||||
},
|
||||
"source": [
|
||||
"# 家庭關係本體\n",
|
||||
"\n",
|
||||
"此範例是 [AI for Beginners Curriculum](http://github.com/microsoft/ai-for-beginners) 的一部分，靈感來自於[這篇部落格文章](https://habr.com/post/270857/)。\n",
|
||||
"\n",
|
||||
"我總是覺得記住家庭中人與人之間的不同關係很困難。在這個範例中，我們將使用一個定義家庭關係的本體，以及實際的家譜樹，來展示如何進行自動推理以找出所有親屬。\n",
|
||||
"\n",
|
||||
"### 獲取家譜樹\n",
|
||||
"\n",
|
||||
"作為範例，我們將使用[羅曼諾夫沙皇家族](https://en.wikipedia.org/wiki/House_of_Romanov)的家譜樹。描述家庭關係最常見的格式是 [GEDCOM](https://en.wikipedia.org/wiki/GEDCOM)。我們將使用 GEDCOM 格式的羅曼諾夫家族樹：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 1,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"0 HEAD\n",
|
||||
"1 CHAR UTF8\n",
|
||||
"1 GEDC\n",
|
||||
"2 VERS 5.5\n",
|
||||
"0 @0@ INDI\n",
|
||||
"1 NAME Mihail Fedorovich /Romanov/\n",
|
||||
"1 SEX M\n",
|
||||
"1 BIRT\n",
|
||||
"2 DATE 1613\n",
|
||||
"1 DEAT \n",
|
||||
"2 DATE 1645\n",
|
||||
"1 FAMS @41@\n",
|
||||
"0 @1@ INDI\n",
|
||||
"1 NAME Evdokija Lukjanovna /Streshneva/\n",
|
||||
"1 SEX F\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"!head -15 data/tsars.ged"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"要使用 GEDCOM 文件，我們可以使用 `python-gedcom` 庫：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 2,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Collecting python-gedcom\n",
|
||||
" Downloading python_gedcom-1.0.0-py2.py3-none-any.whl (35 kB)\n",
|
||||
"Installing collected packages: python-gedcom\n",
|
||||
"Successfully installed python-gedcom-1.0.0\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"import sys\n",
|
||||
"!{sys.executable} -m pip install python-gedcom"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"此程式庫解決了一些與檔案解析相關的技術問題，但仍然提供了對樹中所有個人和家庭的相當低層次的訪問。以下是我們如何解析檔案並顯示所有個人列表的方法：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"from gedcom.parser import Parser\n",
|
||||
"from gedcom.element.individual import IndividualElement\n",
|
||||
"from gedcom.element.family import FamilyElement\n",
|
||||
"g = Parser()\n",
|
||||
"g.parse_file('data/tsars.ged')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {
|
||||
"scrolled": true,
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"[('@0@', ('Mihail Fedorovich', 'Romanov')),\n",
|
||||
" ('@1@', ('Evdokija Lukjanovna', 'Streshneva')),\n",
|
||||
" ('@2@', ('Aleksej Mihajlovich', 'Romanov')),\n",
|
||||
" ('@3@', ('Marija Ilinichna', 'Miloslavskaja')),\n",
|
||||
" ('@4@', ('Natalja Kirillovna', 'Naryshkina')),\n",
|
||||
" ('@5@', ('Marfa Matveevna', 'Apraksina')),\n",
|
||||
" ('@6@', ('Fedor Alekseevich', 'Romanov')),\n",
|
||||
" ('@7@', ('Sofja Aleksevna', 'Romanova')),\n",
|
||||
" ('@8@', ('Ivan V Alekseevich', 'Romanov')),\n",
|
||||
" ('@9@', ('Praskovja Fedorovna', 'Saltykova')),\n",
|
||||
" ('@10@', ('Ekaterina Ivanovna', 'Romanova')),\n",
|
||||
" ('@11@', ('Anna Ivanovna', 'Romanova')),\n",
|
||||
" ('@12@', ('Fridrih Vilgelm', 'Kurlandskij')),\n",
|
||||
" ('@13@', ('Karl Leopold', 'Meklenburg-Shverinskij')),\n",
|
||||
" ('@14@', ('Anna Leopoldovna', 'Meklenburg-Shverinskaja')),\n",
|
||||
" ('@15@', ('Anton Ulrih', 'Braunshvejg-Volfenbjuttelskij')),\n",
|
||||
" ('@16@', ('Ivan VI Antonovich', 'Braunshvejg-Volfenbjuttelskij')),\n",
|
||||
" ('@17@', ('Petr I Alekseevich', 'Romanov')),\n",
|
||||
" ('@18@', ('Evdokija Fedorovna', 'Lopuhina')),\n",
|
||||
" ('@19@', ('Ekaterina I Alekseevna', 'Mihajlova')),\n",
|
||||
" ('@20@', ('Aleksej Petrovich', 'Romanov')),\n",
|
||||
" ('@21@', ('Sharlotta Kristina', 'Braunshvejg-Volfenbjuttelskaja')),\n",
|
||||
" ('@22@', ('Petr II Alekseevich', 'Romanov')),\n",
|
||||
" ('@23@', ('Anna Petrovna', 'Romanova')),\n",
|
||||
" ('@24@', ('Elizaveta Petrovna', 'Romanova')),\n",
|
||||
" ('@25@', ('Karl Fridrih', 'Golshtejn-Gottorpskij')),\n",
|
||||
" ('@26@', ('Petr III Fedorovich', 'Romanov')),\n",
|
||||
" ('@27@', ('Ekaterina II', 'Alekseevna')),\n",
|
||||
" ('@28@', ('Pavel I Petrovich', 'Romanov')),\n",
|
||||
" ('@29@', ('Natalja Alekseevna', 'Gessen-Darmshtadskaja')),\n",
|
||||
" ('@30@', ('Marija Fedorovna', 'Vjurtembergskaja')),\n",
|
||||
" ('@31@', ('Aleksandr I Pavlovich', 'Romanov')),\n",
|
||||
" ('@32@', ('Elizaveta Alekseevna', 'Baden-Durlahskaja')),\n",
|
||||
" ('@33@', ('Nikolaj I Pavlovich', 'Romanov')),\n",
|
||||
" ('@34@', ('Aleksandra Fedorovna', 'Prusskaja')),\n",
|
||||
" ('@35@', ('Aleksandr II Nikolaevich', 'Romanov')),\n",
|
||||
" ('@36@', ('Marija Aleksandrovna', 'Gessenskaja')),\n",
|
||||
" ('@37@', ('Aleksandr III Aleksandrovich', 'Romanov')),\n",
|
||||
" ('@38@', ('Marija Fedorovna', 'Datskaja')),\n",
|
||||
" ('@39@', ('Nikolaj II Aleksandrovich', 'Romanov')),\n",
|
||||
" ('@40@', ('Aleksandra Fedorovna', 'Gessenskaja'))]"
|
||||
]
|
||||
},
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"d = g.get_element_dictionary()\n",
|
||||
"[ (k,v.get_name()) for k,v in d.items() if isinstance(v,IndividualElement)]"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"以下是我們如何獲取有關家庭的信息。請注意，這會給我們一個**標識符**的列表，如果我們想要更清楚的資訊，我們需要將它們轉換為名稱：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"[('@41@', ['@0@', '@1@', '@2@']),\n",
|
||||
" ('@42@', ['@2@', '@3@', '@6@', '@7@', '@8@']),\n",
|
||||
" ('@43@', ['@8@', '@9@', '@10@', '@11@']),\n",
|
||||
" ('@44@', ['@13@', '@10@', '@14@']),\n",
|
||||
" ('@45@', ['@15@', '@14@', '@16@']),\n",
|
||||
" ('@46@', ['@2@', '@4@', '@17@']),\n",
|
||||
" ('@47@', ['@17@', '@18@', '@20@']),\n",
|
||||
" ('@48@', ['@20@', '@21@', '@22@']),\n",
|
||||
" ('@49@', ['@17@', '@19@', '@23@', '@24@']),\n",
|
||||
" ('@50@', ['@25@', '@23@', '@26@']),\n",
|
||||
" ('@51@', ['@26@', '@27@', '@28@']),\n",
|
||||
" ('@52@', ['@28@', '@30@', '@31@', '@33@']),\n",
|
||||
" ('@53@', ['@33@', '@34@', '@35@']),\n",
|
||||
" ('@54@', ['@35@', '@36@', '@37@']),\n",
|
||||
" ('@55@', ['@37@', '@38@', '@39@'])]"
|
||||
]
|
||||
},
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"d = g.get_element_dictionary()\n",
|
||||
"[ (k,[x.get_value() for x in v.get_child_elements()]) for k,v in d.items() if isinstance(v,FamilyElement)]"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"### 獲取家庭本體\n",
|
||||
"\n",
|
||||
"接下來，我們來看看[家庭本體](https://raw.githubusercontent.com/blokhin/genealogical-trees/master/data/header.ttl)，它被定義為一組語義網三元組。這個本體定義了像 `isUncleOf`、`isCousinOf` 等許多關係。所有這些關係都是基於基本謂詞 `isMotherOf`、`isFatherOf`、`isBrotherOf` 和 `isSisterOf` 定義的。我們將使用自動推理來利用該本體推導出所有其他關係。\n",
|
||||
"\n",
|
||||
"以下是 `isAuntOf` 屬性的範例定義，它被定義為 `isSisterOf` 和 `isParentOf` 的組合（*姑姑或阿姨是某人父母的姐妹*）。\n",
|
||||
"\n",
|
||||
"```\n",
|
||||
"fhkb:isAuntOf a owl:ObjectProperty ;\n",
|
||||
" rdfs:domain fhkb:Woman ;\n",
|
||||
" rdfs:range fhkb:Person ;\n",
|
||||
" owl:propertyChainAxiom ( fhkb:isSisterOf fhkb:isParentOf ) .\n",
|
||||
"```\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 6,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"@prefix fhkb: <http://www.example.com/genealogy.owl#> .\n",
|
||||
"@prefix owl: <http://www.w3.org/2002/07/owl#> .\n",
|
||||
"@prefix rdf: <http://www.w3.org/1999/02/22-rdf-syntax-ns#> .\n",
|
||||
"@prefix rdfs: <http://www.w3.org/2000/01/rdf-schema#> .\n",
|
||||
"@prefix xml: <http://www.w3.org/XML/1998/namespace> .\n",
|
||||
"@prefix xsd: <http://www.w3.org/2001/XMLSchema#> .\n",
|
||||
"\n",
|
||||
"<http://www.example.com/genealogy.owl#> a owl:Ontology .\n",
|
||||
"\n",
|
||||
"fhkb:DomainEntity a owl:Class .\n",
|
||||
"\n",
|
||||
"fhkb:Man a owl:Class ;\n",
|
||||
" owl:equivalentClass [ a owl:Class ;\n",
|
||||
" owl:intersectionOf ( fhkb:Person [ a owl:Restriction ;\n",
|
||||
" owl:onProperty fhkb:hasSex ;\n",
|
||||
" owl:someValuesFrom fhkb:Male ] ) ] .\n",
|
||||
"\n",
|
||||
"fhkb:Woman a owl:Class ;\n",
|
||||
" owl:equivalentClass [ a owl:Class ;\n",
|
||||
" owl:intersectionOf ( fhkb:Person [ a owl:Restriction ;\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"!head -20 data/onto.ttl"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"### 建構推理本體\n",
|
||||
"\n",
|
||||
"為了簡化，我們將建立一個本體文件，其中包含家庭本體的原始規則，以及來自我們 GEDCOM 文件的個人事實。我們將逐步處理 GEDCOM 文件，提取有關家庭和個人的信息，並將其轉換為三元組。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 7,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"!cp data/onto.ttl .\n",
|
||||
"\n",
|
||||
"gedcom_dict = g.get_element_dictionary()\n",
|
||||
"individuals, marriages = {}, {}\n",
|
||||
"\n",
|
||||
"def term2id(el):\n",
|
||||
" return \"i\" + el.get_pointer().replace('@', '').lower()\n",
|
||||
"\n",
|
||||
"out = open(\"onto.ttl\",\"a\")\n",
|
||||
"\n",
|
||||
"for k, v in gedcom_dict.items():\n",
|
||||
" if isinstance(v,IndividualElement):\n",
|
||||
" children, siblings = set(), set()\n",
|
||||
" idx = term2id(v)\n",
|
||||
"\n",
|
||||
" title = v.get_name()[0] + \" \" + v.get_name()[1]\n",
|
||||
" title = title.replace('\"', '').replace('[', '').replace(']', '').replace('(', '').replace(')', '').strip()\n",
|
||||
"\n",
|
||||
" own_families = g.get_families(v, 'FAMS')\n",
|
||||
" for fam in own_families:\n",
|
||||
" children |= set(term2id(i) for i in g.get_family_members(fam, \"CHIL\"))\n",
|
||||
"\n",
|
||||
" parent_families = g.get_families(v, 'FAMC')\n",
|
||||
" if len(parent_families):\n",
|
||||
" for member in g.get_family_members(parent_families[0], \"CHIL\"): # NB adoptive families i.e len(parent_families)>1 are not considered (TODO?)\n",
|
||||
" if member.get_pointer() == v.get_pointer():\n",
|
||||
" continue\n",
|
||||
" siblings.add(term2id(member))\n",
|
||||
"\n",
|
||||
" if idx in individuals:\n",
|
||||
" children |= individuals[idx].get('children', set())\n",
|
||||
" siblings |= individuals[idx].get('siblings', set())\n",
|
||||
" individuals[idx] = {'sex': v.get_gender().lower(), 'children': children, 'siblings': siblings, 'title': title}\n",
|
||||
"\n",
|
||||
" elif isinstance(v,FamilyElement):\n",
|
||||
" wife, husb, children = None, None, set()\n",
|
||||
" children = set(term2id(i) for i in g.get_family_members(v, \"CHIL\"))\n",
|
||||
"\n",
|
||||
" try:\n",
|
||||
" wife = g.get_family_members(v, \"WIFE\")[0]\n",
|
||||
" wife = term2id(wife)\n",
|
||||
" if wife in individuals: individuals[wife]['children'] |= children\n",
|
||||
" else: individuals[wife] = {'children': children}\n",
|
||||
" except IndexError: pass\n",
|
||||
" try:\n",
|
||||
" husb = g.get_family_members(v, \"HUSB\")[0]\n",
|
||||
" husb = term2id(husb)\n",
|
||||
" if husb in individuals: individuals[husb]['children'] |= children\n",
|
||||
" else: individuals[husb] = {'children': children}\n",
|
||||
" except IndexError: pass\n",
|
||||
"\n",
|
||||
" if wife and husb: marriages[wife + husb] = (term2id(v), wife, husb)\n",
|
||||
"\n",
|
||||
"for idx, val in individuals.items():\n",
|
||||
" added_terms = ''\n",
|
||||
" if val['sex'] == 'f':\n",
|
||||
" parent_predicate, sibl_predicate = \"isMotherOf\", \"isSisterOf\"\n",
|
||||
" else:\n",
|
||||
" parent_predicate, sibl_predicate = \"isFatherOf\", \"isBrotherOf\"\n",
|
||||
" if len(val['children']):\n",
|
||||
" added_terms += \" ;\\n fhkb:\" + parent_predicate + \" \" + \", \".join([\"fhkb:\" + i for i in val['children']])\n",
|
||||
" if len(val['siblings']):\n",
|
||||
" added_terms += \" ;\\n fhkb:\" + sibl_predicate + \" \" + \", \".join([\"fhkb:\" + i for i in val['siblings']])\n",
|
||||
" out.write(\"fhkb:%s a owl:NamedIndividual, owl:Thing%s ;\\n rdfs:label \\\"%s\\\" .\\n\" % (idx, added_terms, val['title']))\n",
|
||||
"\n",
|
||||
"for k, v in marriages.items():\n",
|
||||
" out.write(\"fhkb:%s a owl:NamedIndividual, owl:Thing ;\\n fhkb:hasFemalePartner fhkb:%s ;\\n fhkb:hasMalePartner fhkb:%s .\\n\" % v)\n",
|
||||
"\n",
|
||||
"out.write(\"[] a owl:AllDifferent ;\\n owl:distinctMembers (\")\n",
|
||||
"for idx in individuals.keys():\n",
|
||||
" out.write(\" fhkb:\" + idx)\n",
|
||||
"for k, v in marriages.items():\n",
|
||||
" out.write(\" fhkb:\" + v[0])\n",
|
||||
"out.write(\" ) .\")\n",
|
||||
"out.close()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 8,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
" fhkb:hasFemalePartner fhkb:i34 ;\n",
|
||||
" fhkb:hasMalePartner fhkb:i33 .\n",
|
||||
"fhkb:i54 a owl:NamedIndividual, owl:Thing ;\n",
|
||||
" fhkb:hasFemalePartner fhkb:i36 ;\n",
|
||||
" fhkb:hasMalePartner fhkb:i35 .\n",
|
||||
"fhkb:i55 a owl:NamedIndividual, owl:Thing ;\n",
|
||||
" fhkb:hasFemalePartner fhkb:i38 ;\n",
|
||||
" fhkb:hasMalePartner fhkb:i37 .\n",
|
||||
"[] a owl:AllDifferent ;\n",
|
||||
" owl:distinctMembers ( fhkb:i0 fhkb:i1 fhkb:i2 fhkb:i3 fhkb:i4 fhkb:i5 fhkb:i6 fhkb:i7 fhkb:i8 fhkb:i9 fhkb:i10 fhkb:i11 fhkb:i12 fhkb:i13 fhkb:i14 fhkb:i15 fhkb:i16 fhkb:i17 fhkb:i18 fhkb:i19 fhkb:i20 fhkb:i21 fhkb:i22 fhkb:i23 fhkb:i24 fhkb:i25 fhkb:i26 fhkb:i27 fhkb:i28 fhkb:i29 fhkb:i30 fhkb:i31 fhkb:i32 fhkb:i33 fhkb:i34 fhkb:i35 fhkb:i36 fhkb:i37 fhkb:i38 fhkb:i39 fhkb:i40 fhkb:i41 fhkb:i42 fhkb:i43 fhkb:i44 fhkb:i45 fhkb:i46 fhkb:i47 fhkb:i48 fhkb:i49 fhkb:i50 fhkb:i51 fhkb:i52 fhkb:i53 fhkb:i54 fhkb:i55 ) ."
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"!tail onto.ttl"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"### 執行推理\n",
|
||||
"\n",
|
||||
"現在我們希望能夠使用這個本體進行推理和查詢。我們將使用 [RDFLib](https://github.com/RDFLib)，這是一個用於以不同格式讀取 RDF 圖、查詢等的庫。\n",
|
||||
"\n",
|
||||
"對於邏輯推理，我們將使用 [OWL-RL](https://github.com/RDFLib/OWL-RL) 庫，該庫允許我們構建 RDF 圖的**閉包**，也就是說，添加所有可以推導出的概念和關係。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 10,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Requirement already satisfied: rdflib in /home/rg/anaconda3/envs/ai4beg/lib/python3.11/site-packages (6.3.2)\n",
|
||||
"Requirement already satisfied: isodate<0.7.0,>=0.6.0 in /home/rg/anaconda3/envs/ai4beg/lib/python3.11/site-packages (from rdflib) (0.6.1)\n",
|
||||
"Requirement already satisfied: pyparsing<4,>=2.1.0 in /home/rg/anaconda3/envs/ai4beg/lib/python3.11/site-packages (from rdflib) (3.0.9)\n",
|
||||
"Requirement already satisfied: six in /home/rg/anaconda3/envs/ai4beg/lib/python3.11/site-packages (from isodate<0.7.0,>=0.6.0->rdflib) (1.16.0)\n",
|
||||
"Collecting git+https://github.com/RDFLib/OWL-RL.git\n",
|
||||
" Cloning https://github.com/RDFLib/OWL-RL.git to /tmp/pip-req-build-lbfzwi3m\n",
|
||||
" Running command git clone --filter=blob:none --quiet https://github.com/RDFLib/OWL-RL.git /tmp/pip-req-build-lbfzwi3m\n",
|
||||
" Resolved https://github.com/RDFLib/OWL-RL.git to commit a77e1791b88b54aace609bc6000aac14c7add4ff\n",
|
||||
" Preparing metadata (setup.py) ... \u001b[?25ldone\n",
|
||||
"\u001b[?25hRequirement already satisfied: rdflib>=6.0.2 in /home/rg/anaconda3/envs/ai4beg/lib/python3.11/site-packages (from owlrl==6.0.2) (6.3.2)\n",
|
||||
"Requirement already satisfied: isodate<0.7.0,>=0.6.0 in /home/rg/anaconda3/envs/ai4beg/lib/python3.11/site-packages (from rdflib>=6.0.2->owlrl==6.0.2) (0.6.1)\n",
|
||||
"Requirement already satisfied: pyparsing<4,>=2.1.0 in /home/rg/anaconda3/envs/ai4beg/lib/python3.11/site-packages (from rdflib>=6.0.2->owlrl==6.0.2) (3.0.9)\n",
|
||||
"Requirement already satisfied: six in /home/rg/anaconda3/envs/ai4beg/lib/python3.11/site-packages (from isodate<0.7.0,>=0.6.0->rdflib>=6.0.2->owlrl==6.0.2) (1.16.0)\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"!{sys.executable} -m pip install rdflib\n",
|
||||
"!{sys.executable} -m pip install git+https://github.com/RDFLib/OWL-RL.git"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"讓我們打開本體文件，看看它包含了多少三元組：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 11,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Triplets found:669\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"import rdflib\n",
|
||||
"from owlrl import DeductiveClosure, OWLRL_Extension\n",
|
||||
"\n",
|
||||
"g = rdflib.Graph()\n",
|
||||
"g.parse(\"onto.ttl\", format=\"turtle\")\n",
|
||||
"\n",
|
||||
"print(\"Triplets found:%d\" % len(g))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在讓我們構建閉包，看看三元組的數量如何增加：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 12,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Triplets after inference:4246\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"DeductiveClosure(OWLRL_Extension).expand(g)\n",
|
||||
"print(\"Triplets after inference:%d\" % len(g))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"### 查詢親屬關係\n",
|
||||
"\n",
|
||||
"現在我們可以查詢圖譜來查看人與人之間的不同關係。我們可以結合使用 **SPARQL** 語言和 `query` 方法。在我們的例子中，讓我們來看看家譜中所有的 **叔叔**：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 13,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Fedor Alekseevich Romanov is uncle of Ekaterina Ivanovna Romanova\n",
|
||||
"Aleksandr I Pavlovich Romanov is uncle of Aleksandr II Nikolaevich Romanov\n",
|
||||
"Fedor Alekseevich Romanov is uncle of Anna Ivanovna Romanova\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"qres = g.query(\n",
|
||||
" \"\"\"SELECT DISTINCT ?aname ?bname\n",
|
||||
" WHERE {\n",
|
||||
" ?a fhkb:isUncleOf ?b .\n",
|
||||
" ?a rdfs:label ?aname .\n",
|
||||
" ?b rdfs:label ?bname .\n",
|
||||
" }\"\"\")\n",
|
||||
"\n",
|
||||
"for row in qres:\n",
|
||||
" print(\"%s is uncle of %s\" % row)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"隨意嘗試不同的家庭關係。例如，您可以查看 `isAncestorOf` 關係，它會遞迴地定義某個人的所有祖先。\n",
|
||||
"\n",
|
||||
"最後，讓我們整理一下！\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 14,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"!rm onto.ttl"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": []
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件使用 AI 翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。儘管我們努力確保翻譯的準確性，但請注意，自動翻譯可能包含錯誤或不準確之處。原始文件的母語版本應被視為權威來源。對於關鍵資訊，建議使用專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或錯誤解釋不承擔責任。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"interpreter": {
|
||||
"hash": "86193a1ab0ba47eac1c69c1756090baa3b420b3eea7d4aafab8b85f8b312f0c5"
|
||||
},
|
||||
"kernelspec": {
|
||||
"display_name": "Python 3.6",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.11.2"
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "6537d5597320e27b6052b4377b8ff8bb",
|
||||
"translation_date": "2025-08-31T10:05:13+00:00",
|
||||
"source_file": "lessons/2-Symbolic/FamilyOntology.ipynb",
|
||||
"language_code": "tw"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 2
|
||||
}
|
||||
|
|
@ -0,0 +1,548 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"collapsed": true
|
||||
},
|
||||
"source": [
|
||||
"## Microsoft 概念圖\n",
|
||||
"\n",
|
||||
"[Microsoft 概念圖](https://concept.research.microsoft.com/) 是一個從網路中挖掘出的大型術語分類，包含概念之間的 `is-a` 關係。\n",
|
||||
"\n",
|
||||
"概念圖提供以下兩種形式：\n",
|
||||
" * 可下載的大型文字檔案\n",
|
||||
" * REST API\n",
|
||||
"\n",
|
||||
"統計數據：\n",
|
||||
" * 5401933 個獨特概念\n",
|
||||
" * 12551613 個獨特實例\n",
|
||||
" * 87603947 個 `is-a` 關係\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 使用網路服務\n",
|
||||
"\n",
|
||||
"網路服務提供不同的調用方式，用於估算某個概念屬於不同群組的可能性。更多資訊可參考[這裡](https://concept.research.microsoft.com/Home/Api)。 \n",
|
||||
"以下是範例 URL：`https://concept.research.microsoft.com/api/Concept/ScoreByProb?instance=microsoft&topK=10`\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 7,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"{'company': 0.6105356614382954,\n",
|
||||
" 'vendor': 0.08858636677518003,\n",
|
||||
" 'client': 0.048239124001183784,\n",
|
||||
" 'firm': 0.045476965571668145,\n",
|
||||
" 'large company': 0.043109401203511886,\n",
|
||||
" 'organization': 0.043010752688172046,\n",
|
||||
" 'corporation': 0.035908059583703265,\n",
|
||||
" 'brand': 0.03383644076156654,\n",
|
||||
" 'software company': 0.027522935779816515,\n",
|
||||
" 'technology company': 0.023774292196902438}"
|
||||
]
|
||||
},
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"import urllib\n",
|
||||
"import json\n",
|
||||
"import ssl\n",
|
||||
"\n",
|
||||
"def http(x):\n",
|
||||
" ssl._create_default_https_context = ssl._create_unverified_context\n",
|
||||
" response = urllib.request.urlopen(x)\n",
|
||||
" data = response.read()\n",
|
||||
" return data.decode('utf-8')\n",
|
||||
"\n",
|
||||
"def query(x):\n",
|
||||
" return json.loads(http(\"https://concept.research.microsoft.com/api/Concept/ScoreByProb?instance={}&topK=10\".format(urllib.parse.quote(x))))\n",
|
||||
"\n",
|
||||
"query('microsoft')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"讓我們嘗試使用父概念對新聞標題進行分類。要獲取新聞標題，我們將使用 [NewsApi.org](http://newsapi.org) 服務。您需要獲取自己的 API 金鑰才能使用該服務——請訪問該網站並註冊免費的開發者計劃。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 20,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"newsapi_key = '<your API key here>'\n",
|
||||
"def get_news(country='us'):\n",
|
||||
" res = json.loads(http(\"https://newsapi.org/v2/top-headlines?country={0}&apiKey={1}\".format(country,newsapi_key)))\n",
|
||||
" return res['articles']\n",
|
||||
"\n",
|
||||
"all_titles = [x['title'] for x in get_news('us')+get_news('gb')]"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 21,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"['Covid-19 Live Updates: Vaccines and Boosters News - The New York Times',\n",
|
||||
" 'Ukrainians Flee Mariupol as Russian Forces Push to Take Port City - The Wall Street Journal',\n",
|
||||
" 'Bond Yields Jump, Stock Futures Rise After Powell Says Fed Is Ready to Be More Aggressive - The Wall Street Journal',\n",
|
||||
" 'Putin critic Alexei Navalny found guilty by Russian court - New York Post ',\n",
|
||||
" \"Supreme Court nominee Ketanji Brown Jackson will face questions at confirmation hearing's second day - CNN\",\n",
|
||||
" '2 teachers killed at Swedish high school, student arrested - ABC News',\n",
|
||||
" 'Clues to Covid-19’s Next Moves Come From Sewers - The Wall Street Journal',\n",
|
||||
" 'Republicans to roll dice by grilling Jackson over child-pornography sentencing decisions | TheHill - The Hill',\n",
|
||||
" '‘Clear sign’ Putin considering using chemical weapons in Ukraine, claims President Biden - The Independent',\n",
|
||||
" 'NASA confirms there are 5,000 planets outside our solar system - Daily Mail',\n",
|
||||
" \"US stocks whipsawed overnight after Fed Chair Powell's remarks - Fox Business\",\n",
|
||||
" \"'We've learned absolutely nothing': Tests could again be in short supply if Covid surges - POLITICO\",\n",
|
||||
" \"Duchess of Cambridge swaps khaki jungle gear for Vampire's Wife dress on Belize trip - Daily Mail\",\n",
|
||||
" 'China searches for victims, flight recorders after first plane crash in 12 years - Reuters',\n",
|
||||
" 'Second superyacht linked to Russian oligarch Abramovich docks in Turkey - Reuters',\n",
|
||||
" 'Live updates: Russia stops talks with Japan over sanctions - The Associated Press - en Español',\n",
|
||||
" 'Powers Remain and Threats Lurk as Women’s Sweet 16 Is Set - The New York Times',\n",
|
||||
" 'Webb Space Telescope Begins Multi-Instrument Alignment - SciTechDaily',\n",
|
||||
" \"UConn vs UCF - NCAA women's tournament second-round highlights - March Madness\",\n",
|
||||
" 'Bucking Republican Trend, Indiana Governor Vetoes Transgender Sports Bill - The New York Times',\n",
|
||||
" \"Maggie Fox dead: Coronation Street and Shameless actress dies after 'sudden accident' - Mirror Online - The Mirror\",\n",
|
||||
" 'China plane crash – live: Search for survivors continues as witness describes moment flight fell from sky - The Independent',\n",
|
||||
" 'Daniel Morgan murder: damning report condemns Met police - The Guardian',\n",
|
||||
" 'What to expect from Rishi Sunak’s Spring Statement - BBC.com',\n",
|
||||
" 'UK and Republic of Ireland in line to host Euro 2028 after no one else bids - The Guardian',\n",
|
||||
" \"Friends beg Vladimir Putin's 'lover' to persuade him to end Ukraine invasion - The Mirror\",\n",
|
||||
" 'Brass Eye’s outtakes show the brutal TV comedy was the tip of an iceberg - The Guardian',\n",
|
||||
" \"Vladimir Putin threatens civilians to break Mariupol's spirit - The Times\",\n",
|
||||
" 'Shell U-turn on Cambo oilfield would threaten green targets, say campaigners - The Guardian',\n",
|
||||
" 'St Helens dog attack: Girl aged 17 months killed at home - BBC',\n",
|
||||
" \"PlayStation to buy 'Assassin's Creed' veteran Jade Raymond's Haven Studios - NME\",\n",
|
||||
" '‘Clear sign’ Putin considering using chemical weapons in Ukraine, claims President Biden - The Independent',\n",
|
||||
" 'NASA confirms there are 5,000 planets outside our solar system - Daily Mail',\n",
|
||||
" 'Nintendo Switch finally has folders • Eurogamer.net - Eurogamer.net',\n",
|
||||
" 'FA to “find a solution” as Liverpool fan group blasts “shambolic” Wembley travel - This Is Anfield',\n",
|
||||
" 'Manchester United transfer news LIVE Erik ten Hag latest and Man Utd manager updates - Manchester Evening News',\n",
|
||||
" 'Inflation raises cost of UK government borrowing in February; crude oil up again – business live - The Guardian',\n",
|
||||
" 'Alexei Navalny: Kremlin critic found guilty of large-scale fraud and contempt of court by Russian court - Sky News',\n",
|
||||
" \"UK prepares to nationalize Russia natural gas giant Gazprom's retail unit - Business Insider\",\n",
|
||||
" 'Zaghari-Ratcliffe: Hunt calls for inquiry into delay over Iran debt payment - The Guardian']"
|
||||
]
|
||||
},
|
||||
"execution_count": 21,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"all_titles"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"首先，我們希望能夠從新聞標題中提取名詞。我們將使用 `TextBlob` 庫來完成這項工作，這使得許多典型的自然語言處理任務變得更加簡單。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 15,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Requirement already satisfied: textblob in c:\\winapp\\miniconda3\\lib\\site-packages (0.17.1)\n",
|
||||
"Requirement already satisfied: nltk>=3.1 in c:\\winapp\\miniconda3\\lib\\site-packages (from textblob) (3.5)\n",
|
||||
"Requirement already satisfied: joblib in c:\\winapp\\miniconda3\\lib\\site-packages (from nltk>=3.1->textblob) (1.0.1)\n",
|
||||
"Requirement already satisfied: regex in c:\\winapp\\miniconda3\\lib\\site-packages (from nltk>=3.1->textblob) (2021.11.10)\n",
|
||||
"Requirement already satisfied: tqdm in c:\\winapp\\miniconda3\\lib\\site-packages (from nltk>=3.1->textblob) (4.61.2)\n",
|
||||
"Requirement already satisfied: click in c:\\winapp\\miniconda3\\lib\\site-packages (from nltk>=3.1->textblob) (8.0.3)\n",
|
||||
"Requirement already satisfied: colorama in c:\\winapp\\miniconda3\\lib\\site-packages (from click->nltk>=3.1->textblob) (0.4.4)\n",
|
||||
"Finished.\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"name": "stderr",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"[nltk_data] Downloading package brown to\n",
|
||||
"[nltk_data] C:\\Users\\dmitryso\\AppData\\Roaming\\nltk_data...\n",
|
||||
"[nltk_data] Package brown is already up-to-date!\n",
|
||||
"[nltk_data] Downloading package punkt to\n",
|
||||
"[nltk_data] C:\\Users\\dmitryso\\AppData\\Roaming\\nltk_data...\n",
|
||||
"[nltk_data] Package punkt is already up-to-date!\n",
|
||||
"[nltk_data] Downloading package wordnet to\n",
|
||||
"[nltk_data] C:\\Users\\dmitryso\\AppData\\Roaming\\nltk_data...\n",
|
||||
"[nltk_data] Package wordnet is already up-to-date!\n",
|
||||
"[nltk_data] Downloading package averaged_perceptron_tagger to\n",
|
||||
"[nltk_data] C:\\Users\\dmitryso\\AppData\\Roaming\\nltk_data...\n",
|
||||
"[nltk_data] Package averaged_perceptron_tagger is already up-to-\n",
|
||||
"[nltk_data] date!\n",
|
||||
"[nltk_data] Downloading package conll2000 to\n",
|
||||
"[nltk_data] C:\\Users\\dmitryso\\AppData\\Roaming\\nltk_data...\n",
|
||||
"[nltk_data] Package conll2000 is already up-to-date!\n",
|
||||
"[nltk_data] Downloading package movie_reviews to\n",
|
||||
"[nltk_data] C:\\Users\\dmitryso\\AppData\\Roaming\\nltk_data...\n",
|
||||
"[nltk_data] Package movie_reviews is already up-to-date!\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"import sys\n",
|
||||
"!{sys.executable} -m pip install textblob\n",
|
||||
"!{sys.executable} -m textblob.download_corpora\n",
|
||||
"from textblob import TextBlob"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 22,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"{'covid-19 live updates': 1,\n",
|
||||
" 'vaccines': 1,\n",
|
||||
" 'boosters': 1,\n",
|
||||
" 'york': 4,\n",
|
||||
" 'ukrainians flee mariupol': 1,\n",
|
||||
" 'forces push': 1,\n",
|
||||
" 'port city': 1,\n",
|
||||
" 'wall street journal': 3,\n",
|
||||
" 'bond yields': 1,\n",
|
||||
" 'futures rise': 1,\n",
|
||||
" 'powell says fed': 1,\n",
|
||||
" 'ready': 1,\n",
|
||||
" 'be': 1,\n",
|
||||
" 'aggressive': 1,\n",
|
||||
" 'putin': 3,\n",
|
||||
" 'alexei navalny': 2,\n",
|
||||
" 'russian': 2,\n",
|
||||
" 'supreme court nominee': 1,\n",
|
||||
" 'ketanji brown jackson': 1,\n",
|
||||
" \"confirmation hearing 's\": 1,\n",
|
||||
" 'cnn': 1,\n",
|
||||
" 'swedish': 1,\n",
|
||||
" 'high school': 1,\n",
|
||||
" 'abc': 1,\n",
|
||||
" 'clues': 1,\n",
|
||||
" 'covid-19': 1,\n",
|
||||
" '’ s': 2,\n",
|
||||
" 'moves': 1,\n",
|
||||
" 'sewers': 1,\n",
|
||||
" 'roll dice': 1,\n",
|
||||
" 'jackson': 1,\n",
|
||||
" 'decisions |': 1,\n",
|
||||
" 'thehill': 1,\n",
|
||||
" 'clear': 2,\n",
|
||||
" 'chemical weapons': 2,\n",
|
||||
" 'ukraine': 3,\n",
|
||||
" 'claims president': 2,\n",
|
||||
" 'biden': 2,\n",
|
||||
" 'nasa': 2,\n",
|
||||
" 'solar system': 2,\n",
|
||||
" 'daily mail': 3,\n",
|
||||
" 'us stocks': 1,\n",
|
||||
" 'fed chair powell': 1,\n",
|
||||
" \"'s remarks\": 1,\n",
|
||||
" 'fox': 1,\n",
|
||||
" \"'we 've\": 1,\n",
|
||||
" 'tests': 1,\n",
|
||||
" 'covid': 1,\n",
|
||||
" 'politico': 1,\n",
|
||||
" 'duchess': 1,\n",
|
||||
" 'cambridge': 1,\n",
|
||||
" 'swaps khaki jungle gear': 1,\n",
|
||||
" 'vampire': 1,\n",
|
||||
" 'wife': 1,\n",
|
||||
" 'belize': 1,\n",
|
||||
" 'china': 2,\n",
|
||||
" 'flight recorders': 1,\n",
|
||||
" 'plane crash': 1,\n",
|
||||
" 'reuters': 2,\n",
|
||||
" 'russian oligarch': 1,\n",
|
||||
" 'abramovich': 1,\n",
|
||||
" 'live': 1,\n",
|
||||
" 'russia': 2,\n",
|
||||
" 'stops talks': 1,\n",
|
||||
" 'japan': 1,\n",
|
||||
" 'español': 1,\n",
|
||||
" 'powers remain': 1,\n",
|
||||
" 'threats lurk': 1,\n",
|
||||
" 'set': 1,\n",
|
||||
" 'webb': 1,\n",
|
||||
" 'telescope begins multi-instrument alignment': 1,\n",
|
||||
" 'scitechdaily': 1,\n",
|
||||
" 'uconn': 1,\n",
|
||||
" 'ucf': 1,\n",
|
||||
" 'ncaa': 1,\n",
|
||||
" \"women 's tournament second-round highlights\": 1,\n",
|
||||
" 'march madness': 1,\n",
|
||||
" 'bucking republican trend': 1,\n",
|
||||
" 'indiana': 1,\n",
|
||||
" 'vetoes transgender': 1,\n",
|
||||
" 'bill': 1,\n",
|
||||
" 'maggie fox': 1,\n",
|
||||
" 'coronation': 1,\n",
|
||||
" 'shameless': 1,\n",
|
||||
" \"'sudden accident\": 1,\n",
|
||||
" 'mirror online': 1,\n",
|
||||
" 'mirror': 2,\n",
|
||||
" 'plane crash –': 1,\n",
|
||||
" 'search': 1,\n",
|
||||
" 'moment flight': 1,\n",
|
||||
" 'daniel morgan': 1,\n",
|
||||
" 'report condemns': 1,\n",
|
||||
" 'met': 1,\n",
|
||||
" 'guardian': 6,\n",
|
||||
" 'rishi sunak': 1,\n",
|
||||
" '’ s spring': 1,\n",
|
||||
" 'statement': 1,\n",
|
||||
" 'bbc.com': 1,\n",
|
||||
" 'uk': 3,\n",
|
||||
" 'ireland': 1,\n",
|
||||
" 'euro': 1,\n",
|
||||
" 'vladimir putin': 2,\n",
|
||||
" \"'s 'lover\": 1,\n",
|
||||
" 'brass eye': 1,\n",
|
||||
" '’ s outtakes': 1,\n",
|
||||
" 'brutal tv comedy': 1,\n",
|
||||
" 'threatens civilians': 1,\n",
|
||||
" 'mariupol': 1,\n",
|
||||
" \"'s spirit\": 1,\n",
|
||||
" 'shell u-turn': 1,\n",
|
||||
" 'cambo': 1,\n",
|
||||
" 'green targets': 1,\n",
|
||||
" 'st helens': 1,\n",
|
||||
" 'dog attack': 1,\n",
|
||||
" 'girl': 1,\n",
|
||||
" 'bbc': 1,\n",
|
||||
" 'playstation': 1,\n",
|
||||
" \"'assassin 's\": 1,\n",
|
||||
" 'creed': 1,\n",
|
||||
" 'jade raymond': 1,\n",
|
||||
" 'haven studios': 1,\n",
|
||||
" 'nme': 1,\n",
|
||||
" 'nintendo switch': 1,\n",
|
||||
" 'folders •': 1,\n",
|
||||
" 'eurogamer.net': 2,\n",
|
||||
" 'fa': 1,\n",
|
||||
" 'solution ”': 1,\n",
|
||||
" 'liverpool': 1,\n",
|
||||
" 'fan group blasts “ shambolic ”': 1,\n",
|
||||
" 'wembley': 1,\n",
|
||||
" 'anfield': 1,\n",
|
||||
" 'manchester': 1,\n",
|
||||
" 'live erik': 1,\n",
|
||||
" 'hag': 1,\n",
|
||||
" 'utd': 1,\n",
|
||||
" 'manager updates': 1,\n",
|
||||
" 'manchester evening': 1,\n",
|
||||
" 'inflation': 1,\n",
|
||||
" 'government borrowing': 1,\n",
|
||||
" 'february': 1,\n",
|
||||
" 'crude oil': 1,\n",
|
||||
" '– business': 1,\n",
|
||||
" 'kremlin': 1,\n",
|
||||
" 'large-scale fraud': 1,\n",
|
||||
" 'sky': 1,\n",
|
||||
" 'natural gas': 1,\n",
|
||||
" 'gazprom': 1,\n",
|
||||
" 'retail unit': 1,\n",
|
||||
" 'insider': 1,\n",
|
||||
" 'zaghari-ratcliffe': 1,\n",
|
||||
" 'hunt': 1,\n",
|
||||
" 'iran': 1,\n",
|
||||
" 'debt payment': 1}"
|
||||
]
|
||||
},
|
||||
"execution_count": 22,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"w = {}\n",
|
||||
"for x in all_titles:\n",
|
||||
" for n in TextBlob(x).noun_phrases:\n",
|
||||
" if n in w:\n",
|
||||
" w[n].append(x)\n",
|
||||
" else:\n",
|
||||
" w[n]=[x]\n",
|
||||
"{ x:len(w[x]) for x in w.keys()}"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們可以看到名詞並未給我們大的主題群組。讓我們用從概念圖中獲得的更一般的術語來替代名詞。這將需要一些時間，因為我們正在為每個名詞短語進行 REST 調用。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 23,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"w = {}\n",
|
||||
"for x in all_titles:\n",
|
||||
" for noun in TextBlob(x).noun_phrases:\n",
|
||||
" terms = query(noun.replace(' ','%20'))\n",
|
||||
" for term in [u for u in terms.keys() if terms[u]>0.1]:\n",
|
||||
" if term in w:\n",
|
||||
" w[term].append(x)\n",
|
||||
" else:\n",
|
||||
" w[term]=[x]"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 24,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"{'city': 9,\n",
|
||||
" 'brand': 4,\n",
|
||||
" 'place': 9,\n",
|
||||
" 'town': 4,\n",
|
||||
" 'factor': 4,\n",
|
||||
" 'film': 4,\n",
|
||||
" 'nation': 11,\n",
|
||||
" 'state': 5,\n",
|
||||
" 'person': 4,\n",
|
||||
" 'organization': 5,\n",
|
||||
" 'publication': 10,\n",
|
||||
" 'market': 5,\n",
|
||||
" 'economy': 4,\n",
|
||||
" 'company': 6,\n",
|
||||
" 'newspaper': 6,\n",
|
||||
" 'relationship': 6}"
|
||||
]
|
||||
},
|
||||
"execution_count": 24,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"{ x:len(w[x]) for x in w.keys() if len(w[x])>3}"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 27,
|
||||
"metadata": {
|
||||
"trusted": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"\n",
|
||||
"ECONOMY:\n",
|
||||
"China searches for victims, flight recorders after first plane crash in 12 years - Reuters\n",
|
||||
"Live updates: Russia stops talks with Japan over sanctions - The Associated Press - en Español\n",
|
||||
"China plane crash – live: Search for survivors continues as witness describes moment flight fell from sky - The Independent\n",
|
||||
"UK prepares to nationalize Russia natural gas giant Gazprom's retail unit - Business Insider\n",
|
||||
"\n",
|
||||
"NATION:\n",
|
||||
"‘Clear sign’ Putin considering using chemical weapons in Ukraine, claims President Biden - The Independent\n",
|
||||
"Duchess of Cambridge swaps khaki jungle gear for Vampire's Wife dress on Belize trip - Daily Mail\n",
|
||||
"China searches for victims, flight recorders after first plane crash in 12 years - Reuters\n",
|
||||
"Live updates: Russia stops talks with Japan over sanctions - The Associated Press - en Español\n",
|
||||
"Live updates: Russia stops talks with Japan over sanctions - The Associated Press - en Español\n",
|
||||
"China plane crash – live: Search for survivors continues as witness describes moment flight fell from sky - The Independent\n",
|
||||
"UK and Republic of Ireland in line to host Euro 2028 after no one else bids - The Guardian\n",
|
||||
"Friends beg Vladimir Putin's 'lover' to persuade him to end Ukraine invasion - The Mirror\n",
|
||||
"‘Clear sign’ Putin considering using chemical weapons in Ukraine, claims President Biden - The Independent\n",
|
||||
"UK prepares to nationalize Russia natural gas giant Gazprom's retail unit - Business Insider\n",
|
||||
"Zaghari-Ratcliffe: Hunt calls for inquiry into delay over Iran debt payment - The Guardian\n",
|
||||
"\n",
|
||||
"PERSON:\n",
|
||||
"‘Clear sign’ Putin considering using chemical weapons in Ukraine, claims President Biden - The Independent\n",
|
||||
"Duchess of Cambridge swaps khaki jungle gear for Vampire's Wife dress on Belize trip - Daily Mail\n",
|
||||
"Second superyacht linked to Russian oligarch Abramovich docks in Turkey - Reuters\n",
|
||||
"‘Clear sign’ Putin considering using chemical weapons in Ukraine, claims President Biden - The Independent\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"print('\\nECONOMY:\\n'+'\\n'.join(w['economy']))\n",
|
||||
"print('\\nNATION:\\n'+'\\n'.join(w['nation']))\n",
|
||||
"print('\\nPERSON:\\n'+'\\n'.join(w['person']))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件已使用 AI 翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。儘管我們致力於提供準確的翻譯，請注意自動翻譯可能包含錯誤或不準確之處。原始文件的母語版本應被視為權威來源。對於關鍵資訊，建議使用專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或錯誤解釋不承擔責任。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"kernelspec": {
|
||||
"display_name": "Python 3.7.4 64-bit (conda)",
|
||||
"metadata": {
|
||||
"interpreter": {
|
||||
"hash": "86193a1ab0ba47eac1c69c1756090baa3b420b3eea7d4aafab8b85f8b312f0c5"
|
||||
}
|
||||
},
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.9.5"
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "4087f998407d06ceb2947016ba4605d0",
|
||||
"translation_date": "2025-08-31T10:06:21+00:00",
|
||||
"source_file": "lessons/2-Symbolic/MSConceptGraph.ipynb",
|
||||
"language_code": "tw"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 2
|
||||
}
|
||||
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
|
|
@ -1,8 +1,8 @@
|
|||
<!--
|
||||
CO_OP_TRANSLATOR_METADATA:
|
||||
{
|
||||
"original_hash": "7336583e4630220c835335da640016db",
|
||||
"translation_date": "2025-08-24T22:10:49+00:00",
|
||||
"original_hash": "ba5d1eb353d20d3e7181066b3c424b99",
|
||||
"translation_date": "2025-08-31T09:18:09+00:00",
|
||||
"source_file": "lessons/3-NeuralNetworks/03-Perceptron/lab/README.md",
|
||||
"language_code": "tw"
|
||||
}
|
||||
|
|
@ -13,19 +13,21 @@ CO_OP_TRANSLATOR_METADATA:
|
|||
|
||||
## 任務
|
||||
|
||||
使用我們在本課中為 MNIST 手寫數字的二元分類所開發的程式碼，創建一個多類別分類器，能夠識別任意數字。計算訓練集和測試集的分類準確率，並輸出混淆矩陣。
|
||||
使用我們在本課中為 MNIST 手寫數字的二元分類所開發的程式碼，創建一個多類別分類器，能夠識別任何數字。計算訓練集和測試集的分類準確率，並輸出混淆矩陣。
|
||||
|
||||
## 提示
|
||||
|
||||
1. 對於每個數字，創建一個二元分類器的數據集，將其定義為「該數字 vs. 其他所有數字」
|
||||
1. 對於每個數字，創建一個二元分類器的數據集，將其設為「該數字 vs. 其他所有數字」
|
||||
1. 訓練 10 個不同的感知器進行二元分類（每個數字一個感知器）
|
||||
1. 定義一個函數來分類輸入的數字
|
||||
|
||||
> **提示**：如果我們將所有 10 個感知器的權重組合成一個矩陣，我們應該能夠通過一次矩陣乘法將所有 10 個感知器應用於輸入數字。然後，只需對輸出應用 `argmax` 操作即可找到最可能的數字。
|
||||
> **提示**：如果我們將所有 10 個感知器的權重組合成一個矩陣，我們應該能夠通過一次矩陣乘法將所有 10 個感知器應用於輸入數字。最可能的數字可以通過對輸出應用 `argmax` 操作來找到。
|
||||
|
||||
## 起始筆記本
|
||||
## 起始 Notebook
|
||||
|
||||
通過打開 [PerceptronMultiClass.ipynb](../../../../../../lessons/3-NeuralNetworks/03-Perceptron/lab/PerceptronMultiClass.ipynb) 開始實驗。
|
||||
通過打開 [PerceptronMultiClass.ipynb](PerceptronMultiClass.ipynb) 開始實驗。
|
||||
|
||||
---
|
||||
|
||||
**免責聲明**：
|
||||
本文件使用 AI 翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。應以原始語言的文件作為權威來源。對於關鍵資訊，建議尋求專業人工翻譯。我們對於因使用此翻譯而引起的任何誤解或錯誤解釋概不負責。
|
||||
本文件已使用 AI 翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。儘管我們致力於提供準確的翻譯，請注意自動翻譯可能包含錯誤或不準確之處。原始文件的母語版本應被視為權威來源。對於關鍵資訊，建議使用專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或錯誤解釋不承擔責任。
|
||||
File diff suppressed because one or more lines are too long
|
|
@ -0,0 +1,183 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 使用我們自己的框架進行 MNIST 數字分類\n",
|
||||
"\n",
|
||||
"來自 [AI 初學者課程](https://github.com/microsoft/ai-for-beginners) 的實驗作業。\n",
|
||||
"\n",
|
||||
"### 讀取數據集\n",
|
||||
"\n",
|
||||
"此程式碼從網路上的存儲庫下載數據集。您也可以手動從 AI 課程存儲庫的 `/data` 目錄中複製數據集。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {
|
||||
"tags": []
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stderr",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
" % Total % Received % Xferd Average Speed Time Time Time Current\n",
|
||||
" Dload Upload Total Spent Left Speed\n",
|
||||
"\n",
|
||||
" 0 0 0 0 0 0 0 0 --:--:-- --:--:-- --:--:-- 0\n",
|
||||
"100 9.9M 100 9.9M 0 0 9.9M 0 0:00:01 --:--:-- 0:00:01 15.8M\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"!rm *.pkl\n",
|
||||
"!wget https://raw.githubusercontent.com/microsoft/AI-For-Beginners/main/data/mnist.pkl.gz\n",
|
||||
"!gzip -d mnist.pkl.gz"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import pickle\n",
|
||||
"with open('mnist.pkl','rb') as f:\n",
|
||||
" MNIST = pickle.load(f)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"labels = MNIST['Train']['Labels']\n",
|
||||
"data = MNIST['Train']['Features']"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"讓我們看看我們擁有的數據形狀：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"(42000, 784)"
|
||||
]
|
||||
},
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"data.shape"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"### 分割資料\n",
|
||||
"\n",
|
||||
"我們將使用 Scikit Learn 將資料分成訓練集和測試集：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 6,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Train samples: 33600, test samples: 8400\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"from sklearn.model_selection import train_test_split\n",
|
||||
"\n",
|
||||
"features_train, features_test, labels_train, labels_test = train_test_split(data,labels,test_size=0.2)\n",
|
||||
"\n",
|
||||
"print(f\"Train samples: {len(features_train)}, test samples: {len(features_test)}\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"### 指示\n",
|
||||
"\n",
|
||||
"1. 從課程中取得框架代碼，並將其貼到此筆記本中，或（更好）貼到一個獨立的 Python 模組中\n",
|
||||
"1. 定義並訓練單層感知器，觀察訓練和驗證的準確率\n",
|
||||
"1. 嘗試了解是否發生了過度擬合，並調整層的參數以提高準確率\n",
|
||||
"1. 重複前面的步驟，訓練兩層和三層感知器。嘗試在層之間使用不同的激活函數進行實驗\n",
|
||||
"1. 嘗試回答以下問題：\n",
|
||||
" - 層間的激活函數是否影響網絡性能？\n",
|
||||
" - 這個任務是否需要兩層或三層的網絡？\n",
|
||||
" - 在訓練網絡時是否遇到任何問題？特別是在層數增加時。\n",
|
||||
" - 網絡的權重在訓練過程中如何表現？你可以繪製權重的最大絕對值與訓練輪次的關係圖來理解這種關係。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": []
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件使用 AI 翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。應以原文文件作為權威來源。對於關鍵資訊，建議尋求專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或誤讀概不負責。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"kernelspec": {
|
||||
"display_name": "Python 3.7.4 64-bit (conda)",
|
||||
"metadata": {
|
||||
"interpreter": {
|
||||
"hash": "86193a1ab0ba47eac1c69c1756090baa3b420b3eea7d4aafab8b85f8b312f0c5"
|
||||
}
|
||||
},
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.9.5"
|
||||
},
|
||||
"orig_nbformat": 2,
|
||||
"coopTranslator": {
|
||||
"original_hash": "6fa055f484eb5d6bdf41166a356d3abf",
|
||||
"translation_date": "2025-08-31T10:12:50+00:00",
|
||||
"source_file": "lessons/3-NeuralNetworks/04-OwnFramework/lab/MyFW_MNIST.ipynb",
|
||||
"language_code": "tw"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 2
|
||||
}
|
||||
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
|
|
@ -0,0 +1,108 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 使用光流檢測手掌移動\n",
|
||||
"\n",
|
||||
"此實驗屬於 [AI 初學者課程](http://aka.ms/ai-beginners)。\n",
|
||||
"\n",
|
||||
"請參考[這段影片](../../../../../../lessons/4-ComputerVision/06-IntroCV/lab/palm-movement.mp4)，影片中一個人的手掌在穩定的背景上向左、向右、向上或向下移動。\n",
|
||||
"\n",
|
||||
"**你的目標**是使用光流技術來判斷影片中哪些部分包含向上、向下、向左或向右的移動。\n",
|
||||
"\n",
|
||||
"首先，按照課程中的說明提取影片幀：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"# Code here"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在，按照講座中描述計算密集光流幀，並將密集光流轉換為極座標：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"# Code here"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"為每個光流幀構建方向的直方圖。直方圖顯示有多少向量落在特定的區間內，並應該將幀中不同方向的運動區分開來。\n",
|
||||
"\n",
|
||||
"> 你可能也會想要將大小低於某個閾值的所有向量歸零。這樣可以消除影片中一些微小的多餘運動，例如眼睛和頭部的移動。\n",
|
||||
"\n",
|
||||
"繪製一些幀的直方圖。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"# Code here"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"看著直方圖，應該很容易判斷移動的方向。你需要選擇那些對應於上/下/左/右方向的區間，並且超過某個閾值的區間。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"# Code here"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"恭喜！如果您已完成以上所有步驟，您已完成實驗！\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件使用 AI 翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。儘管我們努力確保翻譯的準確性，但請注意，自動翻譯可能包含錯誤或不準確之處。原始文件的母語版本應被視為權威來源。對於關鍵資訊，建議使用專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或錯誤解釋不承擔責任。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"language_info": {
|
||||
"name": "python"
|
||||
},
|
||||
"orig_nbformat": 4,
|
||||
"coopTranslator": {
|
||||
"original_hash": "153d9e417e079bf62f8f693002d0deaf",
|
||||
"translation_date": "2025-08-31T09:49:27+00:00",
|
||||
"source_file": "lessons/4-ComputerVision/06-IntroCV/lab/MovementDetection.ipynb",
|
||||
"language_code": "tw"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 2
|
||||
}
|
||||
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
|
|
@ -0,0 +1,577 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 文本分類任務\n",
|
||||
"\n",
|
||||
"如前所述，我們將專注於基於 **AG_NEWS** 數據集的簡單文本分類任務，該任務是將新聞標題分類為以下四個類別之一：世界、體育、商業和科學/技術。\n",
|
||||
"\n",
|
||||
"## 數據集\n",
|
||||
"\n",
|
||||
"此數據集已內建於 [`torchtext`](https://github.com/pytorch/text) 模組中，因此我們可以輕鬆訪問它。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 1,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import torch\n",
|
||||
"import torchtext\n",
|
||||
"import os\n",
|
||||
"import collections\n",
|
||||
"os.makedirs('./data',exist_ok=True)\n",
|
||||
"train_dataset, test_dataset = torchtext.datasets.AG_NEWS(root='./data')\n",
|
||||
"classes = ['World', 'Sports', 'Business', 'Sci/Tech']"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"在這裡，`train_dataset` 和 `test_dataset` 包含分別返回標籤（類別數字）和文本的集合，例如：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 2,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"(3,\n",
|
||||
" \"Wall St. Bears Claw Back Into the Black (Reuters) Reuters - Short-sellers, Wall Street's dwindling\\\\band of ultra-cynics, are seeing green again.\")"
|
||||
]
|
||||
},
|
||||
"execution_count": 2,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"list(train_dataset)[0]"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"那麼，讓我們從數據集中列印出前 10 條新標題：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"**Sci/Tech** -> Wall St. Bears Claw Back Into the Black (Reuters) Reuters - Short-sellers, Wall Street's dwindling\\band of ultra-cynics, are seeing green again.\n",
|
||||
"**Sci/Tech** -> Carlyle Looks Toward Commercial Aerospace (Reuters) Reuters - Private investment firm Carlyle Group,\\which has a reputation for making well-timed and occasionally\\controversial plays in the defense industry, has quietly placed\\its bets on another part of the market.\n",
|
||||
"**Sci/Tech** -> Oil and Economy Cloud Stocks' Outlook (Reuters) Reuters - Soaring crude prices plus worries\\about the economy and the outlook for earnings are expected to\\hang over the stock market next week during the depth of the\\summer doldrums.\n",
|
||||
"**Sci/Tech** -> Iraq Halts Oil Exports from Main Southern Pipeline (Reuters) Reuters - Authorities have halted oil export\\flows from the main pipeline in southern Iraq after\\intelligence showed a rebel militia could strike\\infrastructure, an oil official said on Saturday.\n",
|
||||
"**Sci/Tech** -> Oil prices soar to all-time record, posing new menace to US economy (AFP) AFP - Tearaway world oil prices, toppling records and straining wallets, present a new economic menace barely three months before the US presidential elections.\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"for i,x in zip(range(5),train_dataset):\n",
|
||||
" print(f\"**{classes[x[0]]}** -> {x[1]}\")\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"因為數據集是迭代器，如果我們想多次使用數據，就需要將其轉換為列表：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"train_dataset, test_dataset = torchtext.datasets.AG_NEWS(root='./data')\n",
|
||||
"train_dataset = list(train_dataset)\n",
|
||||
"test_dataset = list(test_dataset)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 分詞\n",
|
||||
"\n",
|
||||
"現在我們需要將文本轉換為可以表示為張量的**數字**。如果我們想要基於單詞的表示，需要完成以下兩件事：\n",
|
||||
"* 使用**分詞器**將文本拆分為**詞元**\n",
|
||||
"* 建立這些詞元的**詞彙表**。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"['he', 'said', 'hello']"
|
||||
]
|
||||
},
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"tokenizer = torchtext.data.utils.get_tokenizer('basic_english')\n",
|
||||
"tokenizer('He said: hello')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"counter = collections.Counter()\n",
|
||||
"for (label, line) in train_dataset:\n",
|
||||
" counter.update(tokenizer(line))\n",
|
||||
"vocab = torchtext.vocab.vocab(counter, min_freq=1)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"使用詞彙，我們可以輕鬆地將標記化的字串編碼為一組數字：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 19,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Vocab size if 95810\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"[599, 3279, 97, 1220, 329, 225, 7368]"
|
||||
]
|
||||
},
|
||||
"execution_count": 19,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"vocab_size = len(vocab)\n",
|
||||
"print(f\"Vocab size if {vocab_size}\")\n",
|
||||
"\n",
|
||||
"stoi = vocab.get_stoi() # dict to convert tokens to indices\n",
|
||||
"\n",
|
||||
"def encode(x):\n",
|
||||
" return [stoi[s] for s in tokenizer(x)]\n",
|
||||
"\n",
|
||||
"encode('I love to play with my words')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 詞袋文字表示法\n",
|
||||
"\n",
|
||||
"由於文字代表意義，有時我們可以僅透過查看個別的文字來理解文本的含義，而不需要考慮它們在句子中的順序。例如，在分類新聞時，像 *weather*、*snow* 這樣的詞可能表明是 *天氣預報*，而像 *stocks*、*dollar* 則可能屬於 *財經新聞*。\n",
|
||||
"\n",
|
||||
"**詞袋** (BoW) 向量表示法是最常用的傳統向量表示法之一。每個詞都與向量索引相關聯，向量元素包含某個詞在特定文件中出現的次數。\n",
|
||||
"\n",
|
||||
" \n",
|
||||
"\n",
|
||||
"> **Note**: 你也可以將 BoW 想像成文本中每個詞的所有獨立一熱編碼向量的總和。\n",
|
||||
"\n",
|
||||
"以下是一個使用 Scikit Learn Python 庫生成詞袋表示法的範例：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"array([[1, 1, 0, 2, 0, 0, 0, 0, 0]], dtype=int64)"
|
||||
]
|
||||
},
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"from sklearn.feature_extraction.text import CountVectorizer\n",
|
||||
"vectorizer = CountVectorizer()\n",
|
||||
"corpus = [\n",
|
||||
" 'I like hot dogs.',\n",
|
||||
" 'The dog ran fast.',\n",
|
||||
" 'Its hot outside.',\n",
|
||||
" ]\n",
|
||||
"vectorizer.fit_transform(corpus)\n",
|
||||
"vectorizer.transform(['My dog likes hot dogs on a hot day.']).toarray()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"要從我們的 AG_NEWS 數據集的向量表示計算詞袋向量，可以使用以下函數：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 20,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"tensor([2., 1., 2., ..., 0., 0., 0.])\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"vocab_size = len(vocab)\n",
|
||||
"\n",
|
||||
"def to_bow(text,bow_vocab_size=vocab_size):\n",
|
||||
" res = torch.zeros(bow_vocab_size,dtype=torch.float32)\n",
|
||||
" for i in encode(text):\n",
|
||||
" if i<bow_vocab_size:\n",
|
||||
" res[i] += 1\n",
|
||||
" return res\n",
|
||||
"\n",
|
||||
"print(to_bow(train_dataset[0][1]))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"> **注意：** 這裡我們使用全域變數 `vocab_size` 來指定詞彙表的預設大小。由於詞彙表的大小通常相當大，我們可以將詞彙表的大小限制為最常出現的詞彙。嘗試降低 `vocab_size` 的值並執行以下程式碼，看看它如何影響準確性。你應該預期準確性會有所下降，但不會太劇烈，以換取更高的效能。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 訓練 BoW 分類器\n",
|
||||
"\n",
|
||||
"現在我們已經學會如何構建文本的詞袋（Bag-of-Words）表示，接下來讓我們基於它來訓練一個分類器。首先，我們需要將數據集轉換為適合訓練的格式，也就是將所有的位置向量表示轉換為詞袋表示。這可以通過將 `bowify` 函數作為 `collate_fn` 參數傳遞給標準的 torch `DataLoader` 來實現：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 21,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"from torch.utils.data import DataLoader\n",
|
||||
"import numpy as np \n",
|
||||
"\n",
|
||||
"# this collate function gets list of batch_size tuples, and needs to \n",
|
||||
"# return a pair of label-feature tensors for the whole minibatch\n",
|
||||
"def bowify(b):\n",
|
||||
" return (\n",
|
||||
" torch.LongTensor([t[0]-1 for t in b]),\n",
|
||||
" torch.stack([to_bow(t[1]) for t in b])\n",
|
||||
" )\n",
|
||||
"\n",
|
||||
"train_loader = DataLoader(train_dataset, batch_size=16, collate_fn=bowify, shuffle=True)\n",
|
||||
"test_loader = DataLoader(test_dataset, batch_size=16, collate_fn=bowify, shuffle=True)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在讓我們定義一個簡單的分類器神經網絡，其中包含一個線性層。輸入向量的大小等於 `vocab_size`，輸出大小對應於類別數量（4）。由於我們正在解決分類任務，最終的激活函數是 `LogSoftmax()`。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 22,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"net = torch.nn.Sequential(torch.nn.Linear(vocab_size,4),torch.nn.LogSoftmax(dim=1))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在我們將定義標準的 PyTorch 訓練循環。由於我們的數據集相當大，出於教學目的，我們將僅訓練一個 epoch，有時甚至少於一個 epoch（通過指定 `epoch_size` 參數可以限制訓練）。我們還會在訓練過程中報告累積的訓練準確率；報告的頻率是通過 `report_freq` 參數指定的。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 24,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"def train_epoch(net,dataloader,lr=0.01,optimizer=None,loss_fn = torch.nn.NLLLoss(),epoch_size=None, report_freq=200):\n",
|
||||
" optimizer = optimizer or torch.optim.Adam(net.parameters(),lr=lr)\n",
|
||||
" net.train()\n",
|
||||
" total_loss,acc,count,i = 0,0,0,0\n",
|
||||
" for labels,features in dataloader:\n",
|
||||
" optimizer.zero_grad()\n",
|
||||
" out = net(features)\n",
|
||||
" loss = loss_fn(out,labels) #cross_entropy(out,labels)\n",
|
||||
" loss.backward()\n",
|
||||
" optimizer.step()\n",
|
||||
" total_loss+=loss\n",
|
||||
" _,predicted = torch.max(out,1)\n",
|
||||
" acc+=(predicted==labels).sum()\n",
|
||||
" count+=len(labels)\n",
|
||||
" i+=1\n",
|
||||
" if i%report_freq==0:\n",
|
||||
" print(f\"{count}: acc={acc.item()/count}\")\n",
|
||||
" if epoch_size and count>epoch_size:\n",
|
||||
" break\n",
|
||||
" return total_loss.item()/count, acc.item()/count"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 25,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"3200: acc=0.8028125\n",
|
||||
"6400: acc=0.8371875\n",
|
||||
"9600: acc=0.8534375\n",
|
||||
"12800: acc=0.85765625\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"(0.026090790722161722, 0.8620069296375267)"
|
||||
]
|
||||
},
|
||||
"execution_count": 25,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"train_epoch(net,train_loader,epoch_size=15000)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## BiGrams、TriGrams 和 N-Grams\n",
|
||||
"\n",
|
||||
"袋裝詞方法的一個限制是，有些詞是多詞表達的一部分。例如，詞語「熱狗」的意思與「熱」和「狗」在其他語境中的意思完全不同。如果我們始終用相同的向量表示「熱」和「狗」，可能會讓模型感到困惑。\n",
|
||||
"\n",
|
||||
"為了解決這個問題，**N-Gram 表示法**通常用於文件分類的方法中，其中每個詞、雙詞或三詞的頻率是訓練分類器的有用特徵。例如，在雙詞表示法中，我們會將所有的詞對加入詞彙表，除了原始詞語之外。\n",
|
||||
"\n",
|
||||
"以下是一個使用 Scikit Learn 生成雙詞袋裝詞表示法的範例：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 26,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Vocabulary:\n",
|
||||
" {'i': 7, 'like': 11, 'hot': 4, 'dogs': 2, 'i like': 8, 'like hot': 12, 'hot dogs': 5, 'the': 16, 'dog': 0, 'ran': 14, 'fast': 3, 'the dog': 17, 'dog ran': 1, 'ran fast': 15, 'its': 9, 'outside': 13, 'its hot': 10, 'hot outside': 6}\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"array([[1, 0, 1, 0, 2, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]],\n",
|
||||
" dtype=int64)"
|
||||
]
|
||||
},
|
||||
"execution_count": 26,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"bigram_vectorizer = CountVectorizer(ngram_range=(1, 2), token_pattern=r'\\b\\w+\\b', min_df=1)\n",
|
||||
"corpus = [\n",
|
||||
" 'I like hot dogs.',\n",
|
||||
" 'The dog ran fast.',\n",
|
||||
" 'Its hot outside.',\n",
|
||||
" ]\n",
|
||||
"bigram_vectorizer.fit_transform(corpus)\n",
|
||||
"print(\"Vocabulary:\\n\",bigram_vectorizer.vocabulary_)\n",
|
||||
"bigram_vectorizer.transform(['My dog likes hot dogs on a hot day.']).toarray()\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"N-gram 方法的主要缺點是詞彙量會以極快的速度增長。實際應用中，我們需要將 N-gram 表示法與一些降維技術結合使用，例如 *嵌入*，我們會在下一單元中討論這些技術。\n",
|
||||
"\n",
|
||||
"為了在我們的 **AG News** 數據集中使用 N-gram 表示法，我們需要構建專門的 ngram 詞彙表：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 27,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Bigram vocabulary length = 1308842\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"counter = collections.Counter()\n",
|
||||
"for (label, line) in train_dataset:\n",
|
||||
" l = tokenizer(line)\n",
|
||||
" counter.update(torchtext.data.utils.ngrams_iterator(l,ngrams=2))\n",
|
||||
" \n",
|
||||
"bi_vocab = torchtext.vocab.vocab(counter, min_freq=1)\n",
|
||||
"\n",
|
||||
"print(\"Bigram vocabulary length = \",len(bi_vocab))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們可以使用上述相同的程式碼來訓練分類器，但這樣做會非常耗費記憶體。在下一個單元中，我們將使用嵌入來訓練二元分類器。\n",
|
||||
"\n",
|
||||
"> **注意:** 你只能保留那些在文本中出現次數超過指定數量的 ngram。這將確保不常見的二元組會被省略，並顯著降低維度。為此，將 `min_freq` 參數設置為更高的值，並觀察詞彙表的長度變化。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 詞頻-逆文檔頻率 TF-IDF\n",
|
||||
"\n",
|
||||
"在 BoW 表示法中，詞的出現次數被均等地加權，無論該詞本身如何。然而，很明顯，像 *a*、*in* 等這些常見詞對分類的作用遠不如專業術語重要。事實上，在大多數 NLP 任務中，有些詞比其他詞更具相關性。\n",
|
||||
"\n",
|
||||
"**TF-IDF** 代表 **詞頻–逆文檔頻率**。它是袋子模型（BoW）的變體，與使用二進制 0/1 值表示詞在文檔中的出現不同，TF-IDF 使用浮點值，該值與詞在語料庫中的出現頻率相關。\n",
|
||||
"\n",
|
||||
"更正式地說，詞 $i$ 在文檔 $j$ 中的權重 $w_{ij}$ 定義為：\n",
|
||||
"$$\n",
|
||||
"w_{ij} = tf_{ij}\\times\\log({N\\over df_i})\n",
|
||||
"$$\n",
|
||||
"其中：\n",
|
||||
"* $tf_{ij}$ 是詞 $i$ 在文檔 $j$ 中的出現次數，即我們之前看到的 BoW 值\n",
|
||||
"* $N$ 是集合中的文檔數量\n",
|
||||
"* $df_i$ 是整個集合中包含詞 $i$ 的文檔數量\n",
|
||||
"\n",
|
||||
"TF-IDF 值 $w_{ij}$ 隨著詞在文檔中出現次數的增加而增加，同時受到包含該詞的文檔數量的影響，這有助於調整某些詞比其他詞出現更頻繁的情況。例如，如果某個詞出現在集合中的*每一個*文檔中，則 $df_i=N$，而 $w_{ij}=0$，這些詞將被完全忽略。\n",
|
||||
"\n",
|
||||
"您可以使用 Scikit Learn 輕鬆地創建文本的 TF-IDF 向量化：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 28,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"array([[0.43381609, 0. , 0.43381609, 0. , 0.65985664,\n",
|
||||
" 0.43381609, 0. , 0. , 0. , 0. ,\n",
|
||||
" 0. , 0. , 0. , 0. , 0. ,\n",
|
||||
" 0. ]])"
|
||||
]
|
||||
},
|
||||
"execution_count": 28,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"from sklearn.feature_extraction.text import TfidfVectorizer\n",
|
||||
"vectorizer = TfidfVectorizer(ngram_range=(1,2))\n",
|
||||
"vectorizer.fit_transform(corpus)\n",
|
||||
"vectorizer.transform(['My dog likes hot dogs on a hot day.']).toarray()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 結論\n",
|
||||
"\n",
|
||||
"然而，儘管 TF-IDF 表示法為不同的詞提供了頻率權重，但它無法表達詞義或順序。正如著名語言學家 J. R. Firth 在1935年所說：「詞語的完整意義總是與上下文相關，任何脫離上下文的意義研究都不應被認真對待。」在課程的後續部分，我們將學習如何通過語言建模從文本中捕捉上下文信息。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件已使用 AI 翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。儘管我們努力確保翻譯的準確性，但請注意，自動翻譯可能包含錯誤或不準確之處。原始文件的母語版本應被視為權威來源。對於關鍵資訊，建議使用專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或錯誤解釋不承擔責任。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"interpreter": {
|
||||
"hash": "16af2a8bbb083ea23e5e41c7f5787656b2ce26968575d8763f2c4b17f9cd711f"
|
||||
},
|
||||
"kernelspec": {
|
||||
"display_name": "Python 3.8.12 ('py38')",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.8.12"
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "7b9040985e748e4e2d4c689892456ad7",
|
||||
"translation_date": "2025-08-31T10:57:29+00:00",
|
||||
"source_file": "lessons/5-NLP/13-TextRep/TextRepresentationPyTorch.ipynb",
|
||||
"language_code": "tw"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 2
|
||||
}
|
||||
|
|
@ -0,0 +1,647 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 文本分類任務\n",
|
||||
"\n",
|
||||
"在本模組中，我們將從一個基於 **[AG_NEWS](http://www.di.unipi.it/~gulli/AG_corpus_of_news_articles.html)** 數據集的簡單文本分類任務開始：我們將把新聞標題分類為以下四個類別之一：世界、體育、商業和科學/技術。\n",
|
||||
"\n",
|
||||
"## 數據集\n",
|
||||
"\n",
|
||||
"為了載入數據集，我們將使用 **[TensorFlow Datasets](https://www.tensorflow.org/datasets)** API。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 1,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import tensorflow as tf\n",
|
||||
"from tensorflow import keras\n",
|
||||
"import tensorflow_datasets as tfds\n",
|
||||
"\n",
|
||||
"# In this tutorial, we will be training a lot of models. In order to use GPU memory cautiously,\n",
|
||||
"# we will set tensorflow option to grow GPU memory allocation when required.\n",
|
||||
"physical_devices = tf.config.list_physical_devices('GPU') \n",
|
||||
"if len(physical_devices)>0:\n",
|
||||
" tf.config.experimental.set_memory_growth(physical_devices[0], True)\n",
|
||||
"\n",
|
||||
"dataset = tfds.load('ag_news_subset')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們現在可以分別使用 `dataset['train']` 和 `dataset['test']` 訪問數據集的訓練部分和測試部分：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Length of train dataset = 120000\n",
|
||||
"Length of test dataset = 7600\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"ds_train = dataset['train']\n",
|
||||
"ds_test = dataset['test']\n",
|
||||
"\n",
|
||||
"print(f\"Length of train dataset = {len(ds_train)}\")\n",
|
||||
"print(f\"Length of test dataset = {len(ds_test)}\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"讓我們列印出資料集中前10個新的標題：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"3 (Sci/Tech) -> b'AMD Debuts Dual-Core Opteron Processor' b'AMD #39;s new dual-core Opteron chip is designed mainly for corporate computing applications, including databases, Web services, and financial transactions.'\n",
|
||||
"1 (Sports) -> b\"Wood's Suspension Upheld (Reuters)\" b'Reuters - Major League Baseball\\\\Monday announced a decision on the appeal filed by Chicago Cubs\\\\pitcher Kerry Wood regarding a suspension stemming from an\\\\incident earlier this season.'\n",
|
||||
"2 (Business) -> b'Bush reform may have blue states seeing red' b'President Bush #39;s quot;revenue-neutral quot; tax reform needs losers to balance its winners, and people claiming the federal deduction for state and local taxes may be in administration planners #39; sights, news reports say.'\n",
|
||||
"3 (Sci/Tech) -> b\"'Halt science decline in schools'\" b'Britain will run out of leading scientists unless science education is improved, says Professor Colin Pillinger.'\n",
|
||||
"1 (Sports) -> b'Gerrard leaves practice' b'London, England (Sports Network) - England midfielder Steven Gerrard injured his groin late in Thursday #39;s training session, but is hopeful he will be ready for Saturday #39;s World Cup qualifier against Austria.'\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"classes = ['World', 'Sports', 'Business', 'Sci/Tech']\n",
|
||||
"\n",
|
||||
"for i,x in zip(range(5),ds_train):\n",
|
||||
" print(f\"{x['label']} ({classes[x['label']]}) -> {x['title']} {x['description']}\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 文本向量化\n",
|
||||
"\n",
|
||||
"現在我們需要將文本轉換為可以表示為張量的**數字**。如果我們想要詞級別的表示，需要完成兩件事：\n",
|
||||
"\n",
|
||||
"* 使用**分詞器**將文本拆分為**詞元**。\n",
|
||||
"* 建立這些詞元的**詞彙表**。\n",
|
||||
"\n",
|
||||
"### 限制詞彙表大小\n",
|
||||
"\n",
|
||||
"在 AG News 數據集的例子中，詞彙表的大小相當大，超過 10 萬個單詞。一般來說，我們不需要那些在文本中很少出現的單詞——只有少數句子會包含它們，模型也無法從中學習。因此，通過向向量化器構造函數傳遞參數，將詞彙表的大小限制為較小的數量是有意義的：\n",
|
||||
"\n",
|
||||
"這兩個步驟都可以使用 **TextVectorization** 層來處理。我們來實例化向量化器對象，然後調用 `adapt` 方法來遍歷所有文本並建立詞彙表：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"vocab_size = 50000\n",
|
||||
"vectorizer = keras.layers.experimental.preprocessing.TextVectorization(max_tokens=vocab_size)\n",
|
||||
"vectorizer.adapt(ds_train.take(500).map(lambda x: x['title']+' '+x['description']))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"> **注意** 我們僅使用整個數據集的一部分來建立詞彙表。這樣做是為了加快執行速度，避免讓您等待太久。然而，我們承擔了一些風險，即整個數據集中的某些詞可能不會被包含在詞彙表中，並在訓練過程中被忽略。因此，使用完整的詞彙表大小並在 `adapt` 過程中遍歷整個數據集應該能提高最終的準確性，但提升幅度不會太大。\n",
|
||||
"\n",
|
||||
"現在我們可以訪問實際的詞彙表：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 6,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"['', '[UNK]', 'the', 'to', 'a', 'in', 'of', 'and', 'on', 'for']\n",
|
||||
"Length of vocabulary: 5335\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"vocab = vectorizer.get_vocabulary()\n",
|
||||
"vocab_size = len(vocab)\n",
|
||||
"print(vocab[:10])\n",
|
||||
"print(f\"Length of vocabulary: {vocab_size}\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"使用向量化器，我們可以輕鬆地將任何文本編碼為一組數字：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<tf.Tensor: shape=(7,), dtype=int64, numpy=array([ 112, 3695, 3, 304, 11, 1041, 1], dtype=int64)>"
|
||||
]
|
||||
},
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"vectorizer('I love to play with my words')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## Bag-of-words 文本表示法\n",
|
||||
"\n",
|
||||
"由於文字具有意義，有時我們僅需查看句子中的個別文字，就能理解一段文字的含義，而不必在意它們的排列順序。例如，在分類新聞時，像 *weather* 和 *snow* 這樣的詞可能表明是 *天氣預報*，而像 *stocks* 和 *dollar* 則可能屬於 *財經新聞*。\n",
|
||||
"\n",
|
||||
"**Bag-of-words** (BoW) 向量表示法是最容易理解的傳統向量表示法。每個詞都與向量索引相關聯，向量中的元素則表示某個詞在特定文件中出現的次數。\n",
|
||||
"\n",
|
||||
" \n",
|
||||
"\n",
|
||||
"> **Note**: 你也可以將 BoW 理解為文本中每個詞的獨熱編碼向量的總和。\n",
|
||||
"\n",
|
||||
"以下是一個使用 Scikit Learn Python 庫生成 Bag-of-words 表示法的範例：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 8,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"array([[1, 1, 0, 2, 0, 0, 0, 0, 0]], dtype=int64)"
|
||||
]
|
||||
},
|
||||
"execution_count": 8,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"from sklearn.feature_extraction.text import CountVectorizer\n",
|
||||
"sc_vectorizer = CountVectorizer()\n",
|
||||
"corpus = [\n",
|
||||
" 'I like hot dogs.',\n",
|
||||
" 'The dog ran fast.',\n",
|
||||
" 'Its hot outside.',\n",
|
||||
" ]\n",
|
||||
"sc_vectorizer.fit_transform(corpus)\n",
|
||||
"sc_vectorizer.transform(['My dog likes hot dogs on a hot day.']).toarray()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們也可以使用我們在上面定義的 Keras 向量化器，將每個單詞編號轉換為一個獨熱編碼，然後將所有這些向量相加：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 9,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"array([0., 5., 0., ..., 0., 0., 0.], dtype=float32)"
|
||||
]
|
||||
},
|
||||
"execution_count": 9,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"def to_bow(text):\n",
|
||||
" return tf.reduce_sum(tf.one_hot(vectorizer(text),vocab_size),axis=0)\n",
|
||||
"\n",
|
||||
"to_bow('My dog likes hot dogs on a hot day.').numpy()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"> **注意**：您可能會驚訝地發現結果與之前的範例不同。原因在於，在 Keras 的範例中，向量的長度對應於詞彙表的大小，而該詞彙表是基於整個 AG News 數據集構建的；而在 Scikit Learn 的範例中，我們是即時從樣本文本中構建詞彙表的。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 訓練 BoW 分類器\n",
|
||||
"\n",
|
||||
"現在我們已經學會如何建立文字的詞袋表示法，接下來讓我們訓練一個使用該表示法的分類器。首先，我們需要將數據集轉換為詞袋表示法。這可以通過以下方式使用 `map` 函數來實現：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 11,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"batch_size = 128\n",
|
||||
"\n",
|
||||
"ds_train_bow = ds_train.map(lambda x: (to_bow(x['title']+x['description']),x['label'])).batch(batch_size)\n",
|
||||
"ds_test_bow = ds_test.map(lambda x: (to_bow(x['title']+x['description']),x['label'])).batch(batch_size)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在讓我們定義一個簡單的分類器神經網絡，其中包含一個線性層。輸入大小為 `vocab_size`，輸出大小對應於類別數量（4）。由於我們正在解決分類任務，最終的激活函數是 **softmax**：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 12,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"938/938 [==============================] - 66s 70ms/step - loss: 0.6144 - acc: 0.8427 - val_loss: 0.4416 - val_acc: 0.8697\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<keras.callbacks.History at 0x20c70a947f0>"
|
||||
]
|
||||
},
|
||||
"execution_count": 12,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model = keras.models.Sequential([\n",
|
||||
" keras.layers.Dense(4,activation='softmax',input_shape=(vocab_size,))\n",
|
||||
"])\n",
|
||||
"model.compile(loss='sparse_categorical_crossentropy',optimizer='adam',metrics=['acc'])\n",
|
||||
"model.fit(ds_train_bow,validation_data=ds_test_bow)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"由於我們有四個類別，超過 80% 的準確率是一個不錯的結果。\n",
|
||||
"\n",
|
||||
"## 將分類器作為一個網絡進行訓練\n",
|
||||
"\n",
|
||||
"由於向量化器本身也是一個 Keras 層，我們可以定義一個包含它的網絡，並進行端到端的訓練。這樣，我們就不需要使用 `map` 來向量化數據集，只需將原始數據集傳遞到網絡的輸入即可。\n",
|
||||
"\n",
|
||||
"> **注意**: 我們仍然需要對數據集應用映射操作，將字典中的字段（例如 `title`、`description` 和 `label`）轉換為元組。然而，在從磁盤加載數據時，我們可以一開始就構建具有所需結構的數據集。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 13,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Model: \"model\"\n",
|
||||
"_________________________________________________________________\n",
|
||||
" Layer (type) Output Shape Param # \n",
|
||||
"=================================================================\n",
|
||||
" input_1 (InputLayer) [(None, 1)] 0 \n",
|
||||
" \n",
|
||||
" text_vectorization (TextVec (None, None) 0 \n",
|
||||
" torization) \n",
|
||||
" \n",
|
||||
" tf.one_hot (TFOpLambda) (None, None, 5335) 0 \n",
|
||||
" \n",
|
||||
" tf.math.reduce_sum (TFOpLam (None, 5335) 0 \n",
|
||||
" bda) \n",
|
||||
" \n",
|
||||
" dense_2 (Dense) (None, 4) 21344 \n",
|
||||
" \n",
|
||||
"=================================================================\n",
|
||||
"Total params: 21,344\n",
|
||||
"Trainable params: 21,344\n",
|
||||
"Non-trainable params: 0\n",
|
||||
"_________________________________________________________________\n",
|
||||
"938/938 [==============================] - 73s 77ms/step - loss: 0.6057 - acc: 0.8414 - val_loss: 0.4202 - val_acc: 0.8736\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<keras.callbacks.History at 0x20c721521f0>"
|
||||
]
|
||||
},
|
||||
"execution_count": 13,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"def extract_text(x):\n",
|
||||
" return x['title']+' '+x['description']\n",
|
||||
"\n",
|
||||
"def tupelize(x):\n",
|
||||
" return (extract_text(x),x['label'])\n",
|
||||
"\n",
|
||||
"inp = keras.Input(shape=(1,),dtype=tf.string)\n",
|
||||
"x = vectorizer(inp)\n",
|
||||
"x = tf.reduce_sum(tf.one_hot(x,vocab_size),axis=1)\n",
|
||||
"out = keras.layers.Dense(4,activation='softmax')(x)\n",
|
||||
"model = keras.models.Model(inp,out)\n",
|
||||
"model.summary()\n",
|
||||
"\n",
|
||||
"model.compile(loss='sparse_categorical_crossentropy',optimizer='adam',metrics=['acc'])\n",
|
||||
"model.fit(ds_train.map(tupelize).batch(batch_size),validation_data=ds_test.map(tupelize).batch(batch_size))\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 二元組、三元組和 n 元組\n",
|
||||
"\n",
|
||||
"詞袋方法的一個限制是，有些詞屬於多詞表達，例如「熱狗」這個詞在不同語境下與「熱」和「狗」的意思完全不同。如果我們始終用相同的向量表示「熱」和「狗」，可能會讓模型感到困惑。\n",
|
||||
"\n",
|
||||
"為了解決這個問題，**n 元組表示法**通常用於文件分類方法中，其中每個詞、雙詞或三詞的頻率是訓練分類器的一個有用特徵。例如，在二元組表示法中，我們會將所有的詞對加入詞彙表中，除了原始詞之外。\n",
|
||||
"\n",
|
||||
"以下是一個使用 Scikit Learn 生成二元組詞袋表示法的示例：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 14,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Vocabulary:\n",
|
||||
" {'i': 7, 'like': 11, 'hot': 4, 'dogs': 2, 'i like': 8, 'like hot': 12, 'hot dogs': 5, 'the': 16, 'dog': 0, 'ran': 14, 'fast': 3, 'the dog': 17, 'dog ran': 1, 'ran fast': 15, 'its': 9, 'outside': 13, 'its hot': 10, 'hot outside': 6}\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"array([[1, 0, 1, 0, 2, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]],\n",
|
||||
" dtype=int64)"
|
||||
]
|
||||
},
|
||||
"execution_count": 14,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"bigram_vectorizer = CountVectorizer(ngram_range=(1, 2), token_pattern=r'\\b\\w+\\b', min_df=1)\n",
|
||||
"corpus = [\n",
|
||||
" 'I like hot dogs.',\n",
|
||||
" 'The dog ran fast.',\n",
|
||||
" 'Its hot outside.',\n",
|
||||
" ]\n",
|
||||
"bigram_vectorizer.fit_transform(corpus)\n",
|
||||
"print(\"Vocabulary:\\n\",bigram_vectorizer.vocabulary_)\n",
|
||||
"bigram_vectorizer.transform(['My dog likes hot dogs on a hot day.']).toarray()\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"n-gram 方法的主要缺點是詞彙量會迅速增長。在實際應用中，我們需要將 n-gram 表示法與降維技術（例如 *嵌入*）結合使用，我們會在下一單元中討論這一點。\n",
|
||||
"\n",
|
||||
"要在我們的 **AG News** 數據集中使用 n-gram 表示法，我們需要將 `ngrams` 參數傳遞給 `TextVectorization` 構造函數。二元語法詞彙的長度**顯著增加**，在我們的例子中，超過了 130 萬個詞元！因此，限制二元語法詞元的數量在合理範圍內是有意義的。\n",
|
||||
"\n",
|
||||
"我們可以使用與上面相同的代碼來訓練分類器，但這樣做會非常耗費記憶體。在下一單元中，我們將使用嵌入來訓練二元語法分類器。與此同時，您可以在這個筆記本中嘗試訓練二元語法分類器，看看是否能獲得更高的準確率。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 自動計算 BoW 向量\n",
|
||||
"\n",
|
||||
"在上述範例中，我們透過手動方式計算 BoW 向量，方法是將個別單字的一次性編碼相加。然而，最新版本的 TensorFlow 允許我們透過在向量化器建構函數中傳入 `output_mode='count` 參數，自動計算 BoW 向量。這使得定義和訓練模型變得更加簡單：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 15,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Training vectorizer\n",
|
||||
"938/938 [==============================] - 7s 7ms/step - loss: 0.5929 - acc: 0.8486 - val_loss: 0.4168 - val_acc: 0.8772\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<keras.callbacks.History at 0x20c725217c0>"
|
||||
]
|
||||
},
|
||||
"execution_count": 15,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model = keras.models.Sequential([\n",
|
||||
" keras.layers.experimental.preprocessing.TextVectorization(max_tokens=vocab_size,output_mode='count'),\n",
|
||||
" keras.layers.Dense(4,input_shape=(vocab_size,), activation='softmax')\n",
|
||||
"])\n",
|
||||
"print(\"Training vectorizer\")\n",
|
||||
"model.layers[0].adapt(ds_train.take(500).map(extract_text))\n",
|
||||
"model.compile(loss='sparse_categorical_crossentropy',optimizer='adam',metrics=['acc'])\n",
|
||||
"model.fit(ds_train.map(tupelize).batch(batch_size),validation_data=ds_test.map(tupelize).batch(batch_size))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 詞頻 - 逆文件頻率 (TF-IDF)\n",
|
||||
"\n",
|
||||
"在 BoW 表示法中，單詞的出現次數無論對於哪個單詞，都是使用相同的技術進行加權。然而，很明顯，像 *a* 和 *in* 這樣的高頻單詞對分類的影響遠不如專業術語。在大多數自然語言處理任務中，有些單詞比其他單詞更為重要。\n",
|
||||
"\n",
|
||||
"**TF-IDF** 是 **詞頻 - 逆文件頻率** 的縮寫。這是一種基於詞袋模型的變體，其中不再使用二元的 0/1 值來表示單詞是否出現在文件中，而是使用與單詞在語料庫中出現頻率相關的浮點值。\n",
|
||||
"\n",
|
||||
"更正式地說，單詞 $i$ 在文件 $j$ 中的權重 $w_{ij}$ 定義為：\n",
|
||||
"$$\n",
|
||||
"w_{ij} = tf_{ij}\\times\\log({N\\over df_i})\n",
|
||||
"$$\n",
|
||||
"其中：\n",
|
||||
"* $tf_{ij}$ 是單詞 $i$ 在文件 $j$ 中出現的次數，也就是我們之前看到的 BoW 值\n",
|
||||
"* $N$ 是集合中的文件總數\n",
|
||||
"* $df_i$ 是整個集合中包含單詞 $i$ 的文件數量\n",
|
||||
"\n",
|
||||
"TF-IDF 值 $w_{ij}$ 與單詞在文件中出現的次數成正比，並且會根據語料庫中包含該單詞的文件數量進行調整，這有助於平衡某些單詞出現頻率較高的情況。例如，如果某個單詞出現在集合中的*每一個*文件中，則 $df_i=N$，而 $w_{ij}=0$，這些詞將被完全忽略。\n",
|
||||
"\n",
|
||||
"您可以使用 Scikit Learn 輕鬆地對文本進行 TF-IDF 向量化：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 16,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"array([[0.43381609, 0. , 0.43381609, 0. , 0.65985664,\n",
|
||||
" 0.43381609, 0. , 0. , 0. , 0. ,\n",
|
||||
" 0. , 0. , 0. , 0. , 0. ,\n",
|
||||
" 0. ]])"
|
||||
]
|
||||
},
|
||||
"execution_count": 16,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"from sklearn.feature_extraction.text import TfidfVectorizer\n",
|
||||
"vectorizer = TfidfVectorizer(ngram_range=(1,2))\n",
|
||||
"vectorizer.fit_transform(corpus)\n",
|
||||
"vectorizer.transform(['My dog likes hot dogs on a hot day.']).toarray()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"在 Keras 中，可以通過傳遞參數 `output_mode='tf-idf'`，讓 `TextVectorization` 層自動計算 TF-IDF 頻率。我們重複上面使用的程式碼，來看看使用 TF-IDF 是否能提高準確性：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 17,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Training vectorizer\n",
|
||||
"938/938 [==============================] - 12s 12ms/step - loss: 0.4197 - acc: 0.8662 - val_loss: 0.3432 - val_acc: 0.8849\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<keras.callbacks.History at 0x20c729dfd30>"
|
||||
]
|
||||
},
|
||||
"execution_count": 17,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model = keras.models.Sequential([\n",
|
||||
" keras.layers.experimental.preprocessing.TextVectorization(max_tokens=vocab_size,output_mode='tf-idf'),\n",
|
||||
" keras.layers.Dense(4,input_shape=(vocab_size,), activation='softmax')\n",
|
||||
"])\n",
|
||||
"print(\"Training vectorizer\")\n",
|
||||
"model.layers[0].adapt(ds_train.take(500).map(extract_text))\n",
|
||||
"model.compile(loss='sparse_categorical_crossentropy',optimizer='adam',metrics=['acc'])\n",
|
||||
"model.fit(ds_train.map(tupelize).batch(batch_size),validation_data=ds_test.map(tupelize).batch(batch_size))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 結論\n",
|
||||
"\n",
|
||||
"儘管 TF-IDF 表示法為不同的詞提供了頻率權重，但它無法表達詞義或順序。正如著名語言學家 J. R. Firth 在 1935 年所說：「詞語的完整意義總是與上下文相關，任何脫離上下文的意義研究都不應被認真對待。」在課程的後續部分，我們將學習如何通過語言建模來捕捉文本中的上下文信息。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件已使用 AI 翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。儘管我們致力於提供準確的翻譯，請注意自動翻譯可能包含錯誤或不準確之處。原始文件的母語版本應被視為權威來源。對於關鍵資訊，建議使用專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或錯誤解釋不承擔責任。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"interpreter": {
|
||||
"hash": "0cb620c6d4b9f7a635928804c26cf22403d89d98d79684e4529119355ee6d5a5"
|
||||
},
|
||||
"kernel_info": {
|
||||
"name": "conda-env-py37_tensorflow-py"
|
||||
},
|
||||
"kernelspec": {
|
||||
"display_name": "py37_tensorflow",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.8.12"
|
||||
},
|
||||
"nteract": {
|
||||
"version": "nteract-front-end@1.0.0"
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "19b43951d55b377a76209c24c1f017e4",
|
||||
"translation_date": "2025-08-31T10:59:55+00:00",
|
||||
"source_file": "lessons/5-NLP/13-TextRep/TextRepresentationTF.ipynb",
|
||||
"language_code": "tw"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 4
|
||||
}
|
||||
|
|
@ -0,0 +1,726 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 嵌入\n",
|
||||
"\n",
|
||||
"在我們之前的例子中，我們操作的是長度為 `vocab_size` 的高維度詞袋向量，並且我們明確地將低維度的位置表示向量轉換為稀疏的獨熱表示（one-hot representation）。這種獨熱表示並不具備記憶效率，此外，每個詞彙彼此之間是獨立處理的，也就是說，獨熱編碼的向量無法表達詞彙之間的任何語義相似性。\n",
|
||||
"\n",
|
||||
"在本單元中，我們將繼續探索 **News AG** 數據集。首先，讓我們載入數據並從之前的筆記本中獲取一些定義。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 1,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Loading dataset...\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"name": "stderr",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"d:\\WORK\\ai-for-beginners\\5-NLP\\14-Embeddings\\data\\train.csv: 29.5MB [00:01, 18.8MB/s] \n",
|
||||
"d:\\WORK\\ai-for-beginners\\5-NLP\\14-Embeddings\\data\\test.csv: 1.86MB [00:00, 11.2MB/s] \n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Building vocab...\n",
|
||||
"Vocab size = 95812\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"import torch\n",
|
||||
"import torchtext\n",
|
||||
"import numpy as np\n",
|
||||
"from torchnlp import *\n",
|
||||
"train_dataset, test_dataset, classes, vocab = load_dataset()\n",
|
||||
"vocab_size = len(vocab)\n",
|
||||
"print(\"Vocab size = \",vocab_size)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 什麼是嵌入？\n",
|
||||
"\n",
|
||||
"**嵌入**的概念是用低維度的密集向量來表示詞語，這些向量以某種方式反映詞語的語義。我們稍後會討論如何構建有意義的詞嵌入，但目前可以將嵌入理解為降低詞向量維度的一種方法。\n",
|
||||
"\n",
|
||||
"嵌入層會將一個詞作為輸入，並生成指定的 `embedding_size` 的輸出向量。從某種意義上說，它與 `Linear` 層非常相似，但嵌入層不需要接收獨熱編碼向量，而是可以直接接收詞的編號作為輸入。\n",
|
||||
"\n",
|
||||
"通過將嵌入層作為我們網絡的第一層，我們可以從詞袋模型切換到 **嵌入袋** 模型。在嵌入袋模型中，我們首先將文本中的每個詞轉換為相應的嵌入，然後對所有這些嵌入計算某種聚合函數，例如 `sum`、`average` 或 `max`。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"我們的分類器神經網絡將以嵌入層開始，接著是聚合層，最後是線性分類器：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 2,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"class EmbedClassifier(torch.nn.Module):\n",
|
||||
" def __init__(self, vocab_size, embed_dim, num_class):\n",
|
||||
" super().__init__()\n",
|
||||
" self.embedding = torch.nn.Embedding(vocab_size, embed_dim)\n",
|
||||
" self.fc = torch.nn.Linear(embed_dim, num_class)\n",
|
||||
"\n",
|
||||
" def forward(self, x):\n",
|
||||
" x = self.embedding(x)\n",
|
||||
" x = torch.mean(x,dim=1)\n",
|
||||
" return self.fc(x)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"### 處理變數序列長度\n",
|
||||
"\n",
|
||||
"由於這種架構的特性，我們需要以特定的方式來建立傳遞給網路的迷你批次。在前一單元中，使用詞袋模型（bag-of-words, BoW）時，迷你批次中的所有 BoW 張量都具有相同的大小 `vocab_size`，無論文本序列的實際長度如何。然而，當我們轉向使用詞嵌入（word embeddings）時，每個文本樣本中的單詞數量會有所不同，而在將這些樣本組合成迷你批次時，我們需要進行一些填充操作。\n",
|
||||
"\n",
|
||||
"這可以通過為數據源提供 `collate_fn` 函數的方式來完成：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"def padify(b):\n",
|
||||
" # b is the list of tuples of length batch_size\n",
|
||||
" # - first element of a tuple = label, \n",
|
||||
" # - second = feature (text sequence)\n",
|
||||
" # build vectorized sequence\n",
|
||||
" v = [encode(x[1]) for x in b]\n",
|
||||
" # first, compute max length of a sequence in this minibatch\n",
|
||||
" l = max(map(len,v))\n",
|
||||
" return ( # tuple of two tensors - labels and features\n",
|
||||
" torch.LongTensor([t[0]-1 for t in b]),\n",
|
||||
" torch.stack([torch.nn.functional.pad(torch.tensor(t),(0,l-len(t)),mode='constant',value=0) for t in v])\n",
|
||||
" )\n",
|
||||
"\n",
|
||||
"train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=16, collate_fn=padify, shuffle=True)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"### 訓練嵌入分類器\n",
|
||||
"\n",
|
||||
"現在我們已經定義了合適的資料加載器，可以使用上一單元中定義的訓練函數來訓練模型：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"3200: acc=0.6415625\n",
|
||||
"6400: acc=0.6865625\n",
|
||||
"9600: acc=0.7103125\n",
|
||||
"12800: acc=0.726953125\n",
|
||||
"16000: acc=0.739375\n",
|
||||
"19200: acc=0.75046875\n",
|
||||
"22400: acc=0.7572321428571429\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"(0.889799795315499, 0.7623160588611644)"
|
||||
]
|
||||
},
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"net = EmbedClassifier(vocab_size,32,len(classes)).to(device)\n",
|
||||
"train_epoch(net,train_loader, lr=1, epoch_size=25000)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"> **注意**：我們在這裡僅訓練 25k 筆記錄（少於一個完整的 epoch）以節省時間，但您可以繼續訓練，撰寫一個函數來訓練多個 epoch，並嘗試調整學習率參數以獲得更高的準確率。您應該能夠達到約 90% 的準確率。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"### EmbeddingBag 層與可變長度序列表示法\n",
|
||||
"\n",
|
||||
"在之前的架構中，我們需要將所有序列填充至相同的長度，以便將它們放入小批量中。這並不是表示可變長度序列的最有效方式——另一種方法是使用 **偏移量** 向量，該向量保存所有序列在一個大型向量中的偏移位置。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"> **Note**: 在上圖中，我們展示的是字符序列，但在我們的例子中，我們處理的是單詞序列。然而，使用偏移量向量表示序列的基本原理是相同的。\n",
|
||||
"\n",
|
||||
"為了使用偏移量表示法，我們使用 [`EmbeddingBag`](https://pytorch.org/docs/stable/generated/torch.nn.EmbeddingBag.html) 層。它類似於 `Embedding`，但它以內容向量和偏移量向量作為輸入，並且還包含一個平均層，可以是 `mean`、`sum` 或 `max`。\n",
|
||||
"\n",
|
||||
"以下是使用 `EmbeddingBag` 的修改後的網絡：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"class EmbedClassifier(torch.nn.Module):\n",
|
||||
" def __init__(self, vocab_size, embed_dim, num_class):\n",
|
||||
" super().__init__()\n",
|
||||
" self.embedding = torch.nn.EmbeddingBag(vocab_size, embed_dim)\n",
|
||||
" self.fc = torch.nn.Linear(embed_dim, num_class)\n",
|
||||
"\n",
|
||||
" def forward(self, text, off):\n",
|
||||
" x = self.embedding(text, off)\n",
|
||||
" return self.fc(x)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"要準備用於訓練的數據集，我們需要提供一個轉換函數來準備偏移向量：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 6,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"def offsetify(b):\n",
|
||||
" # first, compute data tensor from all sequences\n",
|
||||
" x = [torch.tensor(encode(t[1])) for t in b]\n",
|
||||
" # now, compute the offsets by accumulating the tensor of sequence lengths\n",
|
||||
" o = [0] + [len(t) for t in x]\n",
|
||||
" o = torch.tensor(o[:-1]).cumsum(dim=0)\n",
|
||||
" return ( \n",
|
||||
" torch.LongTensor([t[0]-1 for t in b]), # labels\n",
|
||||
" torch.cat(x), # text \n",
|
||||
" o\n",
|
||||
" )\n",
|
||||
"\n",
|
||||
"train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=16, collate_fn=offsetify, shuffle=True)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"注意，與之前所有的例子不同，我們的網絡現在接受兩個參數：數據向量和偏移向量，它們的大小不同。同樣地，我們的數據加載器也提供了3個值而不是2個：文本和偏移向量都作為特徵提供。因此，我們需要稍微調整我們的訓練函數來處理這一點：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"3200: acc=0.6153125\n",
|
||||
"6400: acc=0.6615625\n",
|
||||
"9600: acc=0.6932291666666667\n",
|
||||
"12800: acc=0.715078125\n",
|
||||
"16000: acc=0.7270625\n",
|
||||
"19200: acc=0.7382291666666667\n",
|
||||
"22400: acc=0.7486160714285715\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"(22.771553103007037, 0.7551983365323096)"
|
||||
]
|
||||
},
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"net = EmbedClassifier(vocab_size,32,len(classes)).to(device)\n",
|
||||
"\n",
|
||||
"def train_epoch_emb(net,dataloader,lr=0.01,optimizer=None,loss_fn = torch.nn.CrossEntropyLoss(),epoch_size=None, report_freq=200):\n",
|
||||
" optimizer = optimizer or torch.optim.Adam(net.parameters(),lr=lr)\n",
|
||||
" loss_fn = loss_fn.to(device)\n",
|
||||
" net.train()\n",
|
||||
" total_loss,acc,count,i = 0,0,0,0\n",
|
||||
" for labels,text,off in dataloader:\n",
|
||||
" optimizer.zero_grad()\n",
|
||||
" labels,text,off = labels.to(device), text.to(device), off.to(device)\n",
|
||||
" out = net(text, off)\n",
|
||||
" loss = loss_fn(out,labels) #cross_entropy(out,labels)\n",
|
||||
" loss.backward()\n",
|
||||
" optimizer.step()\n",
|
||||
" total_loss+=loss\n",
|
||||
" _,predicted = torch.max(out,1)\n",
|
||||
" acc+=(predicted==labels).sum()\n",
|
||||
" count+=len(labels)\n",
|
||||
" i+=1\n",
|
||||
" if i%report_freq==0:\n",
|
||||
" print(f\"{count}: acc={acc.item()/count}\")\n",
|
||||
" if epoch_size and count>epoch_size:\n",
|
||||
" break\n",
|
||||
" return total_loss.item()/count, acc.item()/count\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"train_epoch_emb(net,train_loader, lr=4, epoch_size=25000)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 語意嵌入：Word2Vec\n",
|
||||
"\n",
|
||||
"在之前的例子中，模型的嵌入層學會了將文字映射到向量表示，但這種表示並沒有太多語意上的意義。如果能學習到一種向量表示，使得相似的詞或同義詞在某種向量距離（例如歐幾里得距離）上彼此接近，那就更好了。\n",
|
||||
"\n",
|
||||
"為了達到這個目標，我們需要以特定的方式在大量文本上預訓練嵌入模型。最早的語意嵌入訓練方法之一被稱為 [Word2Vec](https://en.wikipedia.org/wiki/Word2vec)。它基於兩種主要架構，用於生成詞的分佈式表示：\n",
|
||||
"\n",
|
||||
" - **連續詞袋模型** (CBoW) — 在這種架構中，我們訓練模型根據周圍的上下文來預測一個詞。給定 ngram $(W_{-2},W_{-1},W_0,W_1,W_2)$，模型的目標是從 $(W_{-2},W_{-1},W_1,W_2)$ 預測 $W_0$。\n",
|
||||
" - **連續跳元模型** (Skip-Gram) 則與 CBoW 相反。模型使用周圍窗口的上下文詞來預測當前詞。\n",
|
||||
"\n",
|
||||
"CBoW 的速度較快，而 Skip-Gram 的速度較慢，但在表示不常見詞方面表現更好。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"要嘗試使用 Google News 數據集預訓練的 Word2Vec 嵌入，我們可以使用 **gensim** 庫。以下是找到與「neural」最相似的詞的示例：\n",
|
||||
"\n",
|
||||
"> **注意:** 當你第一次創建詞向量時，下載它們可能需要一些時間！\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 8,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import gensim.downloader as api\n",
|
||||
"w2v = api.load('word2vec-google-news-300')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 9,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"neuronal -> 0.7804799675941467\n",
|
||||
"neurons -> 0.7326500415802002\n",
|
||||
"neural_circuits -> 0.7252851724624634\n",
|
||||
"neuron -> 0.7174385190010071\n",
|
||||
"cortical -> 0.6941086649894714\n",
|
||||
"brain_circuitry -> 0.6923246383666992\n",
|
||||
"synaptic -> 0.6699118614196777\n",
|
||||
"neural_circuitry -> 0.6638563275337219\n",
|
||||
"neurochemical -> 0.6555314064025879\n",
|
||||
"neuronal_activity -> 0.6531826257705688\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"for w,p in w2v.most_similar('neural'):\n",
|
||||
" print(f\"{w} -> {p}\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們也可以從詞中計算向量嵌入，用於訓練分類模型（為了清楚起見，我們僅顯示向量的前20個組件）：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 10,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"array([ 0.01226807, 0.06225586, 0.10693359, 0.05810547, 0.23828125,\n",
|
||||
" 0.03686523, 0.05151367, -0.20703125, 0.01989746, 0.10058594,\n",
|
||||
" -0.03759766, -0.1015625 , -0.15820312, -0.08105469, -0.0390625 ,\n",
|
||||
" -0.05053711, 0.16015625, 0.2578125 , 0.10058594, -0.25976562],\n",
|
||||
" dtype=float32)"
|
||||
]
|
||||
},
|
||||
"execution_count": 10,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"w2v.word_vec('play')[:20]"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"語義嵌入的優點在於可以操控向量編碼來改變語義。例如，我們可以要求找到一個詞，其向量表示盡可能接近詞 *king* 和 *woman*，並且盡可能遠離詞 *man*：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 10,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"('queen', 0.7118192911148071)"
|
||||
]
|
||||
},
|
||||
"execution_count": 10,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"w2v.most_similar(positive=['king','woman'],negative=['man'])[0]"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"CBoW 和 Skip-Grams 都是「預測型」嵌入，因為它們只考慮局部上下文。Word2Vec 並未利用全局上下文。\n",
|
||||
"\n",
|
||||
"**FastText** 基於 Word2Vec，通過為每個單詞以及單詞內的字符 n-grams 學習向量表示來進一步擴展。在每次訓練步驟中，這些表示的值會被平均成一個向量。雖然這增加了預訓練的計算量，但它使得詞嵌入能夠編碼子詞信息。\n",
|
||||
"\n",
|
||||
"另一種方法，**GloVe**，利用共現矩陣的概念，使用神經方法將共現矩陣分解為更具表達性和非線性的詞向量。\n",
|
||||
"\n",
|
||||
"你可以通過將嵌入模型切換為 FastText 和 GloVe 來試驗這些例子，因為 gensim 支援多種不同的詞嵌入模型。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 在 PyTorch 中使用預訓練的嵌入\n",
|
||||
"\n",
|
||||
"我們可以修改上述範例，將嵌入層中的矩陣預先填入語義嵌入，例如 Word2Vec。我們需要考慮到，預訓練嵌入的詞彙表與我們文本語料庫的詞彙表可能不匹配，因此我們將用隨機值初始化缺失詞彙的權重：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 11,
|
||||
"metadata": {
|
||||
"tags": []
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Embedding size: 300\n",
|
||||
"Populating matrix, this will take some time...Done, found 41080 words, 54732 words missing\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"embed_size = len(w2v.get_vector('hello'))\n",
|
||||
"print(f'Embedding size: {embed_size}')\n",
|
||||
"\n",
|
||||
"net = EmbedClassifier(vocab_size,embed_size,len(classes))\n",
|
||||
"\n",
|
||||
"print('Populating matrix, this will take some time...',end='')\n",
|
||||
"found, not_found = 0,0\n",
|
||||
"for i,w in enumerate(vocab.get_itos()):\n",
|
||||
" try:\n",
|
||||
" net.embedding.weight[i].data = torch.tensor(w2v.get_vector(w))\n",
|
||||
" found+=1\n",
|
||||
" except:\n",
|
||||
" net.embedding.weight[i].data = torch.normal(0.0,1.0,(embed_size,))\n",
|
||||
" not_found+=1\n",
|
||||
"\n",
|
||||
"print(f\"Done, found {found} words, {not_found} words missing\")\n",
|
||||
"net = net.to(device)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在讓我們訓練模型。請注意，由於嵌入層的大小更大，因此參數的數量也大幅增加，訓練模型所需的時間比前一個例子顯著增加。此外，正因如此，如果我們想避免過擬合，可能需要在更多的例子上訓練模型。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 12,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"3200: acc=0.6359375\n",
|
||||
"6400: acc=0.68109375\n",
|
||||
"9600: acc=0.7067708333333333\n",
|
||||
"12800: acc=0.723671875\n",
|
||||
"16000: acc=0.73625\n",
|
||||
"19200: acc=0.7463541666666667\n",
|
||||
"22400: acc=0.7560714285714286\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"(214.1013875559821, 0.7626759436980166)"
|
||||
]
|
||||
},
|
||||
"execution_count": 12,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"train_epoch_emb(net,train_loader, lr=4, epoch_size=25000)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"在我們的情況中，並未看到準確率有顯著提升，這可能是由於詞彙差異較大所致。 \n",
|
||||
"為了解決詞彙差異的問題，我們可以採用以下解決方案之一： \n",
|
||||
"* 重新訓練 word2vec 模型以適應我們的詞彙 \n",
|
||||
"* 使用預訓練的 word2vec 模型的詞彙來載入我們的數據集。在載入數據集時，可以指定使用的詞彙。 \n",
|
||||
"\n",
|
||||
"後者的方法似乎更簡單，尤其是因為 PyTorch 的 `torchtext` 框架內建了對嵌入的支持。例如，我們可以以下列方式實例化基於 GloVe 的詞彙： \n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 14,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stderr",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"100%|█████████▉| 399999/400000 [00:15<00:00, 25411.14it/s]\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"vocab = torchtext.vocab.GloVe(name='6B', dim=50)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"已載入的詞彙表具有以下基本操作：\n",
|
||||
"* `vocab.stoi` 字典允許我們將單詞轉換為其在字典中的索引\n",
|
||||
"* `vocab.itos` 則執行相反的操作——將數字轉換為單詞\n",
|
||||
"* `vocab.vectors` 是嵌入向量的數組，因此要獲取單詞 `s` 的嵌入，我們需要使用 `vocab.vectors[vocab.stoi[s]]`\n",
|
||||
"\n",
|
||||
"以下是一個操作嵌入的範例，用來展示方程式 **kind-man+woman = queen**（我稍微調整了一下係數以使其生效）：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 15,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"'queen'"
|
||||
]
|
||||
},
|
||||
"execution_count": 15,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"# get the vector corresponding to kind-man+woman\n",
|
||||
"qvec = vocab.vectors[vocab.stoi['king']]-vocab.vectors[vocab.stoi['man']]+1.3*vocab.vectors[vocab.stoi['woman']]\n",
|
||||
"# find the index of the closest embedding vector \n",
|
||||
"d = torch.sum((vocab.vectors-qvec)**2,dim=1)\n",
|
||||
"min_idx = torch.argmin(d)\n",
|
||||
"# find the corresponding word\n",
|
||||
"vocab.itos[min_idx]"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"要使用這些嵌入來訓練分類器，我們首先需要使用GloVe詞彙表對我們的數據集進行編碼：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 16,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"def offsetify(b):\n",
|
||||
" # first, compute data tensor from all sequences\n",
|
||||
" x = [torch.tensor(encode(t[1],voc=vocab)) for t in b] # pass the instance of vocab to encode function!\n",
|
||||
" # now, compute the offsets by accumulating the tensor of sequence lengths\n",
|
||||
" o = [0] + [len(t) for t in x]\n",
|
||||
" o = torch.tensor(o[:-1]).cumsum(dim=0)\n",
|
||||
" return ( \n",
|
||||
" torch.LongTensor([t[0]-1 for t in b]), # labels\n",
|
||||
" torch.cat(x), # text \n",
|
||||
" o\n",
|
||||
" )"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"如上所述，所有向量嵌入都存儲在 `vocab.vectors` 矩陣中。這使得通過簡單的複製將這些權重加載到嵌入層的權重中變得非常容易：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 17,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"net = EmbedClassifier(len(vocab),len(vocab.vectors[0]),len(classes))\n",
|
||||
"net.embedding.weight.data = vocab.vectors\n",
|
||||
"net = net.to(device)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在讓我們訓練模型，看看是否能獲得更好的結果：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 18,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"3200: acc=0.6271875\n",
|
||||
"6400: acc=0.68078125\n",
|
||||
"9600: acc=0.7030208333333333\n",
|
||||
"12800: acc=0.71984375\n",
|
||||
"16000: acc=0.7346875\n",
|
||||
"19200: acc=0.7455729166666667\n",
|
||||
"22400: acc=0.7529464285714286\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"(35.53972978646833, 0.7575175943698017)"
|
||||
]
|
||||
},
|
||||
"execution_count": 18,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=16, collate_fn=offsetify, shuffle=True)\n",
|
||||
"train_epoch_emb(net,train_loader, lr=4, epoch_size=25000)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們未看到準確性顯著提高的原因之一是因為我們的數據集中的某些詞語在預訓練的GloVe詞彙表中缺失，因此它們基本上被忽略了。為了克服這一問題，我們可以在我們的數據集上訓練自己的嵌入。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 語境嵌入\n",
|
||||
"\n",
|
||||
"傳統預訓練嵌入表示（例如 Word2Vec）的一個主要限制是詞義消歧的問題。雖然預訓練嵌入可以捕捉到一些單詞在語境中的含義，但每個單詞的所有可能含義都被編碼到同一個嵌入中。這可能會在下游模型中引發問題，因為許多單詞（例如 \"play\"）的含義會根據使用的語境而有所不同。\n",
|
||||
"\n",
|
||||
"例如，單詞 \"play\" 在以下兩個句子中的含義就完全不同：\n",
|
||||
"- 我去劇院看了一場**戲劇**。\n",
|
||||
"- 約翰想和他的朋友們一起**玩**。\n",
|
||||
"\n",
|
||||
"上述的預訓練嵌入將 \"play\" 的這兩種含義表示為相同的嵌入。為了克服這一限制，我們需要基於**語言模型**來構建嵌入，該模型是在大規模文本語料庫上訓練的，並且*知道*單詞如何在不同語境中組合使用。討論語境嵌入超出了本教程的範圍，但我們會在下一單元討論語言模型時回到這個主題。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件使用 AI 翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。應以原文文件作為權威來源。對於關鍵資訊，建議尋求專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或誤讀概不負責。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"interpreter": {
|
||||
"hash": "0cb620c6d4b9f7a635928804c26cf22403d89d98d79684e4529119355ee6d5a5"
|
||||
},
|
||||
"kernelspec": {
|
||||
"display_name": "py37_pytorch",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.8.12"
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "f50b026abce5cf36783a560ea72cb9b1",
|
||||
"translation_date": "2025-08-31T10:54:45+00:00",
|
||||
"source_file": "lessons/5-NLP/14-Embeddings/EmbeddingsPyTorch.ipynb",
|
||||
"language_code": "tw"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 2
|
||||
}
|
||||
|
|
@ -0,0 +1,695 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 嵌入\n",
|
||||
"\n",
|
||||
"在之前的例子中，我們使用了高維度的詞袋向量，其長度為 `vocab_size`，並且我們將低維度的位置表示向量顯式轉換為稀疏的獨熱表示。然而，這種獨熱表示並不具備記憶效率。此外，每個詞都被獨立處理，因此獨熱編碼的向量無法表達詞與詞之間的語義相似性。\n",
|
||||
"\n",
|
||||
"在本單元中，我們將繼續探索 **News AG** 數據集。首先，讓我們載入數據並從上一單元中獲取一些定義。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 2,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import tensorflow as tf\n",
|
||||
"from tensorflow import keras\n",
|
||||
"import tensorflow_datasets as tfds\n",
|
||||
"import numpy as np\n",
|
||||
"\n",
|
||||
"ds_train, ds_test = tfds.load('ag_news_subset').values()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"### 什麼是嵌入？\n",
|
||||
"\n",
|
||||
"**嵌入**的概念是使用低維度的密集向量來表示單詞，這些向量反映了單詞的語義。稍後我們會討論如何構建有意義的單詞嵌入，但目前可以將嵌入理解為一種降低單詞向量維度的方法。\n",
|
||||
"\n",
|
||||
"嵌入層以單詞作為輸入，並生成指定的 `embedding_size` 的輸出向量。某種程度上，它與 `Dense` 層非常相似，但嵌入層不需要以一個 one-hot 編碼向量作為輸入，而是可以直接接受單詞編號。\n",
|
||||
"\n",
|
||||
"通過在網絡中使用嵌入層作為第一層，我們可以從詞袋模型切換到 **嵌入袋** 模型。在嵌入袋模型中，我們首先將文本中的每個單詞轉換為相應的嵌入，然後對所有嵌入進行某種聚合函數計算，例如 `sum`、`average` 或 `max`。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"我們的分類器神經網絡包含以下幾層：\n",
|
||||
"\n",
|
||||
"* `TextVectorization` 層：該層以字符串作為輸入，並生成一個包含標記編號的張量。我們會指定一個合理的詞彙表大小 `vocab_size`，並忽略使用頻率較低的單詞。輸入形狀為 1，輸出形狀為 $n$，因為結果中會有 $n$ 個標記，每個標記包含從 0 到 `vocab_size` 的編號。\n",
|
||||
"* `Embedding` 層：該層接收 $n$ 個編號，並將每個編號縮減為指定長度的密集向量（在我們的例子中為 100）。因此，形狀為 $n$ 的輸入張量將被轉換為 $n\\times 100$ 的張量。\n",
|
||||
"* 聚合層：該層沿第一軸計算該張量的平均值，即計算所有 $n$ 個輸入張量（對應於不同單詞）的平均值。為了實現該層，我們將使用 `Lambda` 層，並向其中傳遞計算平均值的函數。輸出形狀為 100，這將是整個輸入序列的數字表示。\n",
|
||||
"* 最後的 `Dense` 線性分類器。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Model: \"sequential\"\n",
|
||||
"_________________________________________________________________\n",
|
||||
" Layer (type) Output Shape Param # \n",
|
||||
"=================================================================\n",
|
||||
" text_vectorization (TextVec (None, None) 0 \n",
|
||||
" torization) \n",
|
||||
" \n",
|
||||
" embedding (Embedding) (None, None, 100) 3000000 \n",
|
||||
" \n",
|
||||
" lambda (Lambda) (None, 100) 0 \n",
|
||||
" \n",
|
||||
" dense (Dense) (None, 4) 404 \n",
|
||||
" \n",
|
||||
"=================================================================\n",
|
||||
"Total params: 3,000,404\n",
|
||||
"Trainable params: 3,000,404\n",
|
||||
"Non-trainable params: 0\n",
|
||||
"_________________________________________________________________\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"vocab_size = 30000\n",
|
||||
"batch_size = 128\n",
|
||||
"\n",
|
||||
"vectorizer = keras.layers.experimental.preprocessing.TextVectorization(max_tokens=vocab_size,input_shape=(1,))\n",
|
||||
"\n",
|
||||
"model = keras.models.Sequential([\n",
|
||||
" vectorizer, \n",
|
||||
" keras.layers.Embedding(vocab_size,100),\n",
|
||||
" keras.layers.Lambda(lambda x: tf.reduce_mean(x,axis=1)),\n",
|
||||
" keras.layers.Dense(4, activation='softmax')\n",
|
||||
"])\n",
|
||||
"model.summary()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"在 `summary` 列印輸出中，**output shape** 欄位的第一個張量維度 `None` 代表小批量的大小，第二個維度則代表標記序列的長度。小批量中的所有標記序列長度都不相同。我們將在下一節討論如何處理這個問題。\n",
|
||||
"\n",
|
||||
"現在讓我們開始訓練網路：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Training vectorizer\n",
|
||||
"938/938 [==============================] - 20s 20ms/step - loss: 0.7891 - acc: 0.8155 - val_loss: 0.4470 - val_acc: 0.8642\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<keras.callbacks.History at 0x22255515100>"
|
||||
]
|
||||
},
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"def extract_text(x):\n",
|
||||
" return x['title']+' '+x['description']\n",
|
||||
"\n",
|
||||
"def tupelize(x):\n",
|
||||
" return (extract_text(x),x['label'])\n",
|
||||
"\n",
|
||||
"print(\"Training vectorizer\")\n",
|
||||
"vectorizer.adapt(ds_train.take(500).map(extract_text))\n",
|
||||
"\n",
|
||||
"model.compile(loss='sparse_categorical_crossentropy',metrics=['acc'])\n",
|
||||
"model.fit(ds_train.map(tupelize).batch(batch_size),validation_data=ds_test.map(tupelize).batch(batch_size))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"nteract": {
|
||||
"transient": {
|
||||
"deleting": false
|
||||
}
|
||||
}
|
||||
},
|
||||
"source": [
|
||||
"> **注意** 我們正在基於數據的子集構建向量化器。這樣做是為了加快過程，可能會導致我們文本中的某些詞彙未出現在詞彙表中。在這種情況下，這些詞彙將被忽略，這可能會導致準確性略有下降。然而，在現實生活中，文本的子集通常能夠提供良好的詞彙估計。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"### 處理變化的序列大小\n",
|
||||
"\n",
|
||||
"讓我們來了解如何在小批量中進行訓練。在上面的例子中，輸入張量的維度為 1，我們使用 128 長度的小批量，因此張量的實際大小為 $128 \\times 1$。然而，每個句子中的標記數量是不同的。如果我們將 `TextVectorization` 層應用於單一輸入，返回的標記數量會有所不同，這取決於文本如何被分詞：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"tf.Tensor([ 1 45], shape=(2,), dtype=int64)\n",
|
||||
"tf.Tensor([ 112 1271 1 3 1747 158], shape=(6,), dtype=int64)\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"print(vectorizer('Hello, world!'))\n",
|
||||
"print(vectorizer('I am glad to meet you!'))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"然而，當我們將向量化器應用於多個序列時，它必須生成一個矩形形狀的張量，因此會用 PAD 標記（在我們的情況下是零）填充未使用的元素：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 6,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<tf.Tensor: shape=(2, 6), dtype=int64, numpy=\n",
|
||||
"array([[ 1, 45, 0, 0, 0, 0],\n",
|
||||
" [ 112, 1271, 1, 3, 1747, 158]], dtype=int64)>"
|
||||
]
|
||||
},
|
||||
"execution_count": 6,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"vectorizer(['Hello, world!','I am glad to meet you!'])"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"在這裡我們可以看到嵌入：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"array([[[ 1.53059261e-02, 6.80514947e-02, 3.14026810e-02, ...,\n",
|
||||
" -8.92002955e-02, 1.52911525e-04, -5.65562584e-02],\n",
|
||||
" [ 2.57456154e-01, 2.79364467e-01, -2.03605562e-01, ...,\n",
|
||||
" -2.07474351e-01, 8.31158683e-02, -2.03911960e-01],\n",
|
||||
" [ 3.98201384e-02, -8.03454965e-03, 2.39790026e-02, ...,\n",
|
||||
" -7.18549127e-04, 2.66963355e-02, -4.30646613e-02],\n",
|
||||
" [ 3.98201384e-02, -8.03454965e-03, 2.39790026e-02, ...,\n",
|
||||
" -7.18549127e-04, 2.66963355e-02, -4.30646613e-02],\n",
|
||||
" [ 3.98201384e-02, -8.03454965e-03, 2.39790026e-02, ...,\n",
|
||||
" -7.18549127e-04, 2.66963355e-02, -4.30646613e-02],\n",
|
||||
" [ 3.98201384e-02, -8.03454965e-03, 2.39790026e-02, ...,\n",
|
||||
" -7.18549127e-04, 2.66963355e-02, -4.30646613e-02]],\n",
|
||||
"\n",
|
||||
" [[ 1.89674050e-01, 2.61548996e-01, -3.67433839e-02, ...,\n",
|
||||
" -2.07366899e-01, -1.05442435e-01, -2.36952081e-01],\n",
|
||||
" [ 6.16133213e-02, 1.80511594e-01, 9.77298319e-02, ...,\n",
|
||||
" -5.46628237e-02, -1.07340455e-01, -1.06589928e-01],\n",
|
||||
" [ 1.53059261e-02, 6.80514947e-02, 3.14026810e-02, ...,\n",
|
||||
" -8.92002955e-02, 1.52911525e-04, -5.65562584e-02],\n",
|
||||
" [-4.84890305e-02, -8.41715634e-02, 1.51529670e-01, ...,\n",
|
||||
" 1.28192469e-01, -7.77286515e-02, 1.26041949e-01],\n",
|
||||
" [-4.17212099e-02, -5.60694858e-02, 4.08860669e-02, ...,\n",
|
||||
" 8.70475471e-02, 8.92383084e-02, 1.67974353e-01],\n",
|
||||
" [ 2.85779923e-01, 4.57767487e-01, 4.52292450e-02, ...,\n",
|
||||
" -1.97419018e-01, -2.04659685e-01, -2.79758364e-01]]],\n",
|
||||
" dtype=float32)"
|
||||
]
|
||||
},
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model.layers[1](vectorizer(['Hello, world!','I am glad to meet you!'])).numpy()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"> **注意**：為了減少填充的數量，在某些情況下，將數據集中所有序列按長度（更準確地說是按標記數量）遞增排序是有意義的。這將確保每個小批次包含相似長度的序列。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 語意嵌入：Word2Vec\n",
|
||||
"\n",
|
||||
"在我們之前的例子中，嵌入層學習將單詞映射到向量表示，但這些表示並不具有語意上的意義。如果能學習一種向量表示，使得相似的單詞或同義詞在某種向量距離（例如歐幾里得距離）上彼此接近，那就更好了。\n",
|
||||
"\n",
|
||||
"為了實現這一點，我們需要使用像 [Word2Vec](https://en.wikipedia.org/wiki/Word2vec) 這樣的技術，在大量文本集合上預訓練嵌入模型。它基於兩種主要架構來生成單詞的分佈式表示：\n",
|
||||
"\n",
|
||||
" - **連續詞袋模型** (CBoW)，我們訓練模型根據周圍的上下文來預測一個單詞。給定 ngram $(W_{-2},W_{-1},W_0,W_1,W_2)$，模型的目標是根據 $(W_{-2},W_{-1},W_1,W_2)$ 預測 $W_0$。\n",
|
||||
" - **連續跳字模型** (Skip-Gram) 則與 CBoW 相反。模型使用周圍窗口中的上下文單詞來預測當前單詞。\n",
|
||||
"\n",
|
||||
"CBoW 的速度較快，而 Skip-Gram 雖然速度較慢，但在表示不常見單詞方面效果更佳。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"為了試驗基於 Google News 數據集預訓練的 Word2Vec 嵌入，我們可以使用 **gensim** 庫。以下是找到與「neural」最相似的單詞的示例。\n",
|
||||
"\n",
|
||||
"> **注意:** 當您首次創建單詞向量時，下載它們可能需要一些時間！\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 8,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import gensim.downloader as api\n",
|
||||
"w2v = api.load('word2vec-google-news-300')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 12,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"neuronal -> 0.7804799675941467\n",
|
||||
"neurons -> 0.7326500415802002\n",
|
||||
"neural_circuits -> 0.7252851724624634\n",
|
||||
"neuron -> 0.7174385190010071\n",
|
||||
"cortical -> 0.6941086649894714\n",
|
||||
"brain_circuitry -> 0.6923246383666992\n",
|
||||
"synaptic -> 0.6699118614196777\n",
|
||||
"neural_circuitry -> 0.6638563275337219\n",
|
||||
"neurochemical -> 0.6555314064025879\n",
|
||||
"neuronal_activity -> 0.6531826257705688\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"for w,p in w2v.most_similar('neural'):\n",
|
||||
" print(f\"{w} -> {p}\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們還可以從該詞中提取向量嵌入，用於訓練分類模型。該嵌入有300個組件，但這裡為了清楚起見，我們僅顯示向量的前20個組件：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 13,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"array([ 0.01226807, 0.06225586, 0.10693359, 0.05810547, 0.23828125,\n",
|
||||
" 0.03686523, 0.05151367, -0.20703125, 0.01989746, 0.10058594,\n",
|
||||
" -0.03759766, -0.1015625 , -0.15820312, -0.08105469, -0.0390625 ,\n",
|
||||
" -0.05053711, 0.16015625, 0.2578125 , 0.10058594, -0.25976562],\n",
|
||||
" dtype=float32)"
|
||||
]
|
||||
},
|
||||
"execution_count": 13,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"w2v['play'][:20]"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"語義嵌入的偉大之處在於您可以根據語義操控向量編碼。例如，我們可以要求找到一個其向量表示盡可能接近*國王*和*女人*，並且盡可能遠離*男人*的詞：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 14,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"('queen', 0.7118192911148071)"
|
||||
]
|
||||
},
|
||||
"execution_count": 14,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"w2v.most_similar(positive=['king','woman'],negative=['man'])[0]"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"tags": []
|
||||
},
|
||||
"source": [
|
||||
"上面的例子使用了一些內部的 GenSym 魔法，但其底層邏輯其實相當簡單。關於嵌入的一個有趣之處是，你可以對嵌入向量執行正常的向量運算，而這將反映在詞語**意義**上的操作。上面的例子可以用向量運算來表達：我們計算出對應於 **KING-MAN+WOMAN** 的向量（對應詞語的向量表示執行 `+` 和 `-` 運算），然後在字典中找到最接近該向量的詞語：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 15,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"'queen'"
|
||||
]
|
||||
},
|
||||
"execution_count": 15,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"# get the vector corresponding to kind-man+woman\n",
|
||||
"qvec = w2v['king']-1.7*w2v['man']+1.7*w2v['woman']\n",
|
||||
"# find the index of the closest embedding vector \n",
|
||||
"d = np.sum((w2v.vectors-qvec)**2,axis=1)\n",
|
||||
"min_idx = np.argmin(d)\n",
|
||||
"# find the corresponding word\n",
|
||||
"w2v.index_to_key[min_idx]"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"> **NOTE**: 我們需要在 *man* 和 *woman* 向量中添加一個小係數 - 試著移除它們看看會發生什麼。\n",
|
||||
"\n",
|
||||
"為了找到最接近的向量，我們使用 TensorFlow 的機制來計算我們的向量與詞彙表中所有向量之間的距離向量，然後使用 `argmin` 找到最小距離的詞的索引。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"雖然 Word2Vec 看起來是一種表達詞語語義的好方法，但它有許多缺點，包括以下幾點：\n",
|
||||
"\n",
|
||||
"* CBoW 和 skip-gram 模型都是**預測型嵌入**，它們只考慮局部上下文。Word2Vec 無法利用全局上下文。\n",
|
||||
"* Word2Vec 沒有考慮到詞語的**形態學**，也就是說，詞語的意義可能取決於詞的不同部分，例如詞根。\n",
|
||||
"\n",
|
||||
"**FastText** 試圖克服第二個限制，並在 Word2Vec 的基礎上進一步改進，通過學習每個詞以及詞內字符 n-grams 的向量表示來實現。在每次訓練步驟中，這些表示的值會被平均為一個向量。雖然這為預訓練增加了大量的計算，但它使詞嵌入能夠編碼子詞資訊。\n",
|
||||
"\n",
|
||||
"另一種方法，**GloVe**，採用了與 Word2Vec 不同的詞嵌入方法，基於詞-上下文矩陣的分解。首先，它構建了一個大型矩陣，記錄詞語在不同上下文中的出現次數，然後嘗試以降低維度的方式表示該矩陣，從而最小化重建損失。\n",
|
||||
"\n",
|
||||
"gensim 庫支持這些詞嵌入方法，您可以通過更改上方的模型加載代碼來嘗試使用它們。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 在 Keras 中使用預訓練的嵌入\n",
|
||||
"\n",
|
||||
"我們可以修改上述範例，將嵌入層中的矩陣預先填入語義嵌入，例如 Word2Vec。預訓練嵌入的詞彙表和文本語料庫的詞彙表可能不匹配，因此我們需要選擇其中一個。在這裡，我們探討兩種可能的選擇：使用分詞器的詞彙表，或使用 Word2Vec 嵌入的詞彙表。\n",
|
||||
"\n",
|
||||
"### 使用分詞器詞彙表\n",
|
||||
"\n",
|
||||
"當使用分詞器的詞彙表時，詞彙表中的某些單詞會有對應的 Word2Vec 嵌入，而某些則會缺失。假設我們的詞彙表大小為 `vocab_size`，而 Word2Vec 嵌入向量的長度為 `embed_size`，那麼嵌入層將由一個形狀為 `vocab_size`$\\times$`embed_size` 的權重矩陣表示。我們將通過遍歷詞彙表來填充這個矩陣：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 9,
|
||||
"metadata": {
|
||||
"tags": []
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Embedding size: 300\n",
|
||||
"Populating matrix, this will take some time...Done, found 4551 words, 784 words missing\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"embed_size = len(w2v.get_vector('hello'))\n",
|
||||
"print(f'Embedding size: {embed_size}')\n",
|
||||
"\n",
|
||||
"vocab = vectorizer.get_vocabulary()\n",
|
||||
"W = np.zeros((vocab_size,embed_size))\n",
|
||||
"print('Populating matrix, this will take some time...',end='')\n",
|
||||
"found, not_found = 0,0\n",
|
||||
"for i,w in enumerate(vocab):\n",
|
||||
" try:\n",
|
||||
" W[i] = w2v.get_vector(w)\n",
|
||||
" found+=1\n",
|
||||
" except:\n",
|
||||
" # W[i] = np.random.normal(0.0,0.3,size=(embed_size,))\n",
|
||||
" not_found+=1\n",
|
||||
"\n",
|
||||
"print(f\"Done, found {found} words, {not_found} words missing\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"對於不在 Word2Vec 詞彙表中的單詞，我們可以選擇將它們設為零向量，或者生成一個隨機向量。\n",
|
||||
"\n",
|
||||
"現在我們可以定義一個帶有預訓練權重的嵌入層：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 10,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"emb = keras.layers.Embedding(vocab_size,embed_size,weights=[W],trainable=False)\n",
|
||||
"model = keras.models.Sequential([\n",
|
||||
" vectorizer, emb,\n",
|
||||
" keras.layers.Lambda(lambda x: tf.reduce_mean(x,axis=1)),\n",
|
||||
" keras.layers.Dense(4, activation='softmax')\n",
|
||||
"])"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": []
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 11,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"938/938 [==============================] - 10s 10ms/step - loss: 1.1075 - acc: 0.7822 - val_loss: 0.9134 - val_acc: 0.8175\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<keras.callbacks.History at 0x2220226ef10>"
|
||||
]
|
||||
},
|
||||
"execution_count": 11,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model.compile(loss='sparse_categorical_crossentropy',metrics=['acc'])\n",
|
||||
"model.fit(ds_train.map(tupelize).batch(batch_size),\n",
|
||||
" validation_data=ds_test.map(tupelize).batch(batch_size))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"> **注意**：請注意，我們在建立 `Embedding` 時設置了 `trainable=False`，這表示我們不會重新訓練 Embedding 層。這可能會導致準確率稍微降低，但能加快訓練速度。\n",
|
||||
"\n",
|
||||
"### 使用嵌入詞彙表\n",
|
||||
"\n",
|
||||
"之前方法的一個問題是，TextVectorization 和 Embedding 使用的詞彙表不同。為了解決這個問題，我們可以採用以下其中一種解決方案：\n",
|
||||
"* 重新訓練 Word2Vec 模型以適配我們的詞彙表。\n",
|
||||
"* 使用預訓練 Word2Vec 模型的詞彙表來載入我們的數據集。在載入數據集時，可以指定使用的詞彙表。\n",
|
||||
"\n",
|
||||
"第二種方法看起來更簡單，因此我們來實現它。首先，我們將使用從 Word2Vec 嵌入中提取的指定詞彙表來建立一個 `TextVectorization` 層：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 12,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"vocab = list(w2v.vocab.keys())\n",
|
||||
"vectorizer = keras.layers.experimental.preprocessing.TextVectorization(input_shape=(1,))\n",
|
||||
"vectorizer.set_vocabulary(vocab)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"gensim 的詞嵌入庫包含一個方便的函數 `get_keras_embeddings`，它會自動為您創建相應的 Keras 嵌入層。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 13,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Epoch 1/5\n",
|
||||
"938/938 [==============================] - 20s 14ms/step - loss: 1.3377 - acc: 0.4978 - val_loss: 1.2995 - val_acc: 0.5647\n",
|
||||
"Epoch 2/5\n",
|
||||
"938/938 [==============================] - 10s 10ms/step - loss: 1.2587 - acc: 0.5722 - val_loss: 1.2339 - val_acc: 0.5842\n",
|
||||
"Epoch 3/5\n",
|
||||
"938/938 [==============================] - 10s 10ms/step - loss: 1.1980 - acc: 0.5884 - val_loss: 1.1826 - val_acc: 0.5954\n",
|
||||
"Epoch 4/5\n",
|
||||
"938/938 [==============================] - 12s 13ms/step - loss: 1.1503 - acc: 0.6002 - val_loss: 1.1417 - val_acc: 0.6018\n",
|
||||
"Epoch 5/5\n",
|
||||
"938/938 [==============================] - 11s 12ms/step - loss: 1.1120 - acc: 0.6097 - val_loss: 1.1083 - val_acc: 0.6104\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<keras.callbacks.History at 0x2220ccb81c0>"
|
||||
]
|
||||
},
|
||||
"execution_count": 13,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model = keras.models.Sequential([\n",
|
||||
" vectorizer, \n",
|
||||
" w2v.get_keras_embedding(train_embeddings=False),\n",
|
||||
" keras.layers.Lambda(lambda x: tf.reduce_mean(x,axis=1)),\n",
|
||||
" keras.layers.Dense(4, activation='softmax')\n",
|
||||
"])\n",
|
||||
"model.compile(loss='sparse_categorical_crossentropy',metrics=['acc'])\n",
|
||||
"model.fit(ds_train.map(tupelize).batch(128),validation_data=ds_test.map(tupelize).batch(128),epochs=5)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們未能看到更高準確率的原因之一是因為我們的數據集中有些詞在預訓練的 GloVe 詞彙表中缺失，因此它們實際上被忽略了。為了解決這個問題，我們可以基於我們的數據集訓練我們自己的詞嵌入。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 語境嵌入\n",
|
||||
"\n",
|
||||
"傳統的預訓練嵌入表示（例如 Word2Vec）的一個主要限制是，儘管它們能捕捉到某些詞語的含義，但無法區分不同的含義。這可能會在下游模型中引發問題。\n",
|
||||
"\n",
|
||||
"例如，\"play\" 這個詞在以下兩個句子中有不同的含義：\n",
|
||||
"- 我去劇院看了一場**戲劇**。\n",
|
||||
"- 約翰想和他的朋友們一起**玩**。\n",
|
||||
"\n",
|
||||
"我們之前提到的預訓練嵌入會將 \"play\" 的兩種含義表示為相同的嵌入。為了克服這一限制，我們需要基於**語言模型**來構建嵌入。語言模型是在大量文本語料庫上訓練的，並且*了解*詞語如何在不同的語境中組合使用。討論語境嵌入超出了本教程的範圍，但我們會在下一單元討論語言模型時回到這個主題。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件使用 AI 翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。應以原文文件作為權威來源。對於關鍵資訊，建議尋求專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或誤讀概不負責。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"interpreter": {
|
||||
"hash": "0cb620c6d4b9f7a635928804c26cf22403d89d98d79684e4529119355ee6d5a5"
|
||||
},
|
||||
"kernel_info": {
|
||||
"name": "conda-env-py37_tensorflow-py"
|
||||
},
|
||||
"kernelspec": {
|
||||
"display_name": "py37_tensorflow",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.8.12"
|
||||
},
|
||||
"nteract": {
|
||||
"version": "nteract-front-end@1.0.0"
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "b859482be7f61d1eadc2c6a2720a37e4",
|
||||
"translation_date": "2025-08-31T10:51:43+00:00",
|
||||
"source_file": "lessons/5-NLP/14-Embeddings/EmbeddingsTF.ipynb",
|
||||
"language_code": "tw"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 4
|
||||
}
|
||||
|
|
@ -0,0 +1,576 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"id": "NXTSugt6ieXh"
|
||||
},
|
||||
"source": [
|
||||
"## 訓練 CBoW 模型\n",
|
||||
"\n",
|
||||
"此筆記本是 [AI for Beginners Curriculum](http://aka.ms/ai-beginners) 的一部分\n",
|
||||
"\n",
|
||||
"在這個範例中，我們將探討如何訓練 CBoW 語言模型來獲得我們自己的 Word2Vec 嵌入空間。我們將使用 AG News 數據集作為文本來源。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"import torch\n",
|
||||
"import torchtext\n",
|
||||
"import os\n",
|
||||
"import collections\n",
|
||||
"import builtins\n",
|
||||
"import random\n",
|
||||
"import numpy as np"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "q-UiiJUKaxHj"
|
||||
},
|
||||
"execution_count": null,
|
||||
"outputs": []
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "TFbR8CZaTZ1q"
|
||||
},
|
||||
"execution_count": null,
|
||||
"outputs": []
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"source": [
|
||||
"首先，我們來載入數據集並定義分詞器和詞彙表。我們將 `vocab_size` 設置為 5000，以稍微限制計算量。\n"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "HIwC7lI5T-ov"
|
||||
}
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"def load_dataset(ngrams = 1, min_freq = 1, vocab_size = 5000 , lines_cnt = 500):\n",
|
||||
" tokenizer = torchtext.data.utils.get_tokenizer('basic_english')\n",
|
||||
" print(\"Loading dataset...\")\n",
|
||||
" test_dataset, train_dataset = torchtext.datasets.AG_NEWS(root='./data')\n",
|
||||
" train_dataset = list(train_dataset)\n",
|
||||
" test_dataset = list(test_dataset)\n",
|
||||
" classes = ['World', 'Sports', 'Business', 'Sci/Tech']\n",
|
||||
" print('Building vocab...')\n",
|
||||
" counter = collections.Counter()\n",
|
||||
" for i, (_, line) in enumerate(train_dataset):\n",
|
||||
" counter.update(torchtext.data.utils.ngrams_iterator(tokenizer(line),ngrams=ngrams))\n",
|
||||
" if i == lines_cnt:\n",
|
||||
" break\n",
|
||||
" vocab = torchtext.vocab.Vocab(collections.Counter(dict(counter.most_common(vocab_size))), min_freq=min_freq)\n",
|
||||
" return train_dataset, test_dataset, classes, vocab, tokenizer"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "wdZuygtgiuLG"
|
||||
},
|
||||
"execution_count": null,
|
||||
"outputs": []
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"train_dataset, test_dataset, _, vocab, tokenizer = load_dataset()"
|
||||
],
|
||||
"metadata": {
|
||||
"colab": {
|
||||
"base_uri": "https://localhost:8080/"
|
||||
},
|
||||
"id": "4d1nU1gsivGu",
|
||||
"outputId": "949fe272-ae0e-49f5-c373-6703458b3a74"
|
||||
},
|
||||
"execution_count": null,
|
||||
"outputs": [
|
||||
{
|
||||
"output_type": "stream",
|
||||
"name": "stdout",
|
||||
"text": [
|
||||
"Loading dataset...\n",
|
||||
"Building vocab...\n"
|
||||
]
|
||||
}
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"def encode(x, vocabulary, tokenizer = tokenizer):\n",
|
||||
" return [vocabulary[s] for s in tokenizer(x)]"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "1XDYNhG8ToFV"
|
||||
},
|
||||
"execution_count": null,
|
||||
"outputs": []
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"id": "LIlQk6_PaHVY"
|
||||
},
|
||||
"source": [
|
||||
"## CBoW 模型\n",
|
||||
"\n",
|
||||
"CBoW 學習根據 $2N$ 個鄰近的詞來預測一個詞。例如，當 $N=1$ 時，我們可以從句子 *I like to train networks* 中得到以下配對：(like,I)、(I, like)、(to, like)、(like,to)、(train,to)、(to, train)、(networks, train)、(train,networks)。在這裡，第一個詞是作為輸入的鄰近詞，第二個詞是我們要預測的詞。\n",
|
||||
"\n",
|
||||
"為了構建一個用於預測下一個詞的網絡，我們需要提供鄰近詞作為輸入，並獲得詞的編號作為輸出。CBoW 網絡的架構如下：\n",
|
||||
"\n",
|
||||
"* 輸入詞會通過嵌入層。這個嵌入層就是我們的 Word2Vec 嵌入，因此我們會將其單獨定義為 `embedder` 變數。在這個例子中，我們將使用嵌入大小 = 30，儘管你可能想要嘗試更高的維度（真實的 Word2Vec 嵌入大小為 300）。\n",
|
||||
"* 嵌入向量接著會傳遞到一個線性層，該層將預測輸出詞。因此它有 `vocab_size` 個神經元。\n",
|
||||
"\n",
|
||||
"對於輸出，如果我們使用 `CrossEntropyLoss` 作為損失函數，我們只需要提供詞的編號作為預期結果，而不需要使用 one-hot 編碼。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"vocab_size = len(vocab)\n",
|
||||
"\n",
|
||||
"embedder = torch.nn.Embedding(num_embeddings = vocab_size, embedding_dim = 30)\n",
|
||||
"model = torch.nn.Sequential(\n",
|
||||
" embedder,\n",
|
||||
" torch.nn.Linear(in_features = 30, out_features = vocab_size),\n",
|
||||
")\n",
|
||||
"\n",
|
||||
"print(model)"
|
||||
],
|
||||
"metadata": {
|
||||
"colab": {
|
||||
"base_uri": "https://localhost:8080/"
|
||||
},
|
||||
"id": "akKTcKQKkfl2",
|
||||
"outputId": "da687e3e-a8ec-4c1a-e456-ab8cd6ac7dad"
|
||||
},
|
||||
"execution_count": null,
|
||||
"outputs": [
|
||||
{
|
||||
"output_type": "stream",
|
||||
"name": "stdout",
|
||||
"text": [
|
||||
"Sequential(\n",
|
||||
" (0): Embedding(5002, 30)\n",
|
||||
" (1): Linear(in_features=30, out_features=5002, bias=True)\n",
|
||||
")\n"
|
||||
]
|
||||
}
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"id": "Nud6jgGPaHVa"
|
||||
},
|
||||
"source": [
|
||||
"## 準備訓練數據\n",
|
||||
"\n",
|
||||
"現在讓我們編寫主要函數，用於從文本中計算 CBoW 單詞對。這個函數將允許我們指定窗口大小，並返回一組配對——輸入詞和輸出詞。請注意，這個函數可以用於單詞，也可以用於向量/張量——這將允許我們在將文本傳遞給 `to_cbow` 函數之前對其進行編碼。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {
|
||||
"colab": {
|
||||
"base_uri": "https://localhost:8080/"
|
||||
},
|
||||
"id": "x-dsXygOieXn",
|
||||
"outputId": "c2218280-e540-40ba-9546-efe48d0d714f"
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"output_type": "stream",
|
||||
"name": "stdout",
|
||||
"text": [
|
||||
"[['like', 'I'], ['to', 'I'], ['I', 'like'], ['to', 'like'], ['train', 'like'], ['I', 'to'], ['like', 'to'], ['train', 'to'], ['networks', 'to'], ['like', 'train'], ['to', 'train'], ['networks', 'train'], ['to', 'networks'], ['train', 'networks']]\n",
|
||||
"[[232, 172], [5, 172], [172, 232], [5, 232], [0, 232], [172, 5], [232, 5], [0, 5], [1202, 5], [232, 0], [5, 0], [1202, 0], [5, 1202], [0, 1202]]\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"def to_cbow(sent,window_size=2):\n",
|
||||
" res = []\n",
|
||||
" for i,x in enumerate(sent):\n",
|
||||
" for j in range(max(0,i-window_size),min(i+window_size+1,len(sent))):\n",
|
||||
" if i!=j:\n",
|
||||
" res.append([sent[j],x])\n",
|
||||
" return res\n",
|
||||
"\n",
|
||||
"print(to_cbow(['I','like','to','train','networks']))\n",
|
||||
"print(to_cbow(encode('I like to train networks', vocab)))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"id": "XVaaDLjaaHVb"
|
||||
},
|
||||
"source": [
|
||||
"讓我們準備訓練數據集。我們將遍歷所有新聞，調用 `to_cbow` 來獲取單詞對列表，並將這些對添加到 `X` 和 `Y` 中。為了節省時間，我們只考慮前 10k 條新聞項目——如果你有更多時間等待並希望獲得更好的嵌入，可以輕鬆移除此限制 :)\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {
|
||||
"id": "54b-Gd9TieXo"
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"X = []\n",
|
||||
"Y = []\n",
|
||||
"for i, x in zip(range(10000), train_dataset):\n",
|
||||
" for w1, w2 in to_cbow(encode(x[1], vocab), window_size = 5):\n",
|
||||
" X.append(w1)\n",
|
||||
" Y.append(w2)\n",
|
||||
"\n",
|
||||
"X = torch.tensor(X)\n",
|
||||
"Y = torch.tensor(Y)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"source": [
|
||||
"我們還將把該數據轉換為一個數據集，並創建數據加載器：\n"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "cwWy0PzXWhN5"
|
||||
}
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"class SimpleIterableDataset(torch.utils.data.IterableDataset):\n",
|
||||
" def __init__(self, X, Y):\n",
|
||||
" super(SimpleIterableDataset).__init__()\n",
|
||||
" self.data = []\n",
|
||||
" for i in range(len(X)):\n",
|
||||
" self.data.append( (Y[i], X[i]) )\n",
|
||||
" random.shuffle(self.data)\n",
|
||||
"\n",
|
||||
" def __iter__(self):\n",
|
||||
" return iter(self.data)"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "mfoAcGPFZU8p"
|
||||
},
|
||||
"execution_count": null,
|
||||
"outputs": []
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"id": "e4NQ_-5waHVc"
|
||||
},
|
||||
"source": [
|
||||
"我們還將把該數據轉換為一個數據集，並創建數據加載器：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {
|
||||
"id": "AbLUcojlieXo"
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"ds = SimpleIterableDataset(X, Y)\n",
|
||||
"dl = torch.utils.data.DataLoader(ds, batch_size = 256)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"id": "pKQr7sXeaHVc"
|
||||
},
|
||||
"source": [
|
||||
"現在讓我們進行實際訓練。我們將使用 `SGD` 優化器，並設定相當高的學習率。你也可以嘗試使用其他優化器，例如 `Adam`。我們將從訓練 10 個世代開始——如果你想要更低的損失，可以重新執行此單元格。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"def train_epoch(net, dataloader, lr = 0.01, optimizer = None, loss_fn = torch.nn.CrossEntropyLoss(), epochs = None, report_freq = 1):\n",
|
||||
" optimizer = optimizer or torch.optim.Adam(net.parameters(), lr = lr)\n",
|
||||
" loss_fn = loss_fn.to(device)\n",
|
||||
" net.train()\n",
|
||||
"\n",
|
||||
" for i in range(epochs):\n",
|
||||
" total_loss, j = 0, 0, \n",
|
||||
" for labels, features in dataloader:\n",
|
||||
" optimizer.zero_grad()\n",
|
||||
" features, labels = features.to(device), labels.to(device)\n",
|
||||
" out = net(features)\n",
|
||||
" loss = loss_fn(out, labels)\n",
|
||||
" loss.backward()\n",
|
||||
" optimizer.step()\n",
|
||||
" total_loss += loss\n",
|
||||
" j += 1\n",
|
||||
" if i % report_freq == 0:\n",
|
||||
" print(f\"Epoch: {i+1}: loss={total_loss.item()/j}\")\n",
|
||||
"\n",
|
||||
" return total_loss.item()/j"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "HeeCYKr_KF1w"
|
||||
},
|
||||
"execution_count": null,
|
||||
"outputs": []
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"train_epoch(net = model, dataloader = dl, optimizer = torch.optim.SGD(model.parameters(), lr = 0.1), loss_fn = torch.nn.CrossEntropyLoss(), epochs = 10)"
|
||||
],
|
||||
"metadata": {
|
||||
"colab": {
|
||||
"base_uri": "https://localhost:8080/"
|
||||
},
|
||||
"id": "KVgwGtDHgDlT",
|
||||
"outputId": "2447833f-f0e3-4566-c33d-addbfe2f451d"
|
||||
},
|
||||
"execution_count": null,
|
||||
"outputs": [
|
||||
{
|
||||
"output_type": "stream",
|
||||
"name": "stdout",
|
||||
"text": [
|
||||
"Epoch: 1: loss=5.664632366860172\n",
|
||||
"Epoch: 2: loss=5.632101973960962\n",
|
||||
"Epoch: 3: loss=5.610399051405015\n",
|
||||
"Epoch: 4: loss=5.594621561080262\n",
|
||||
"Epoch: 5: loss=5.582538017415446\n",
|
||||
"Epoch: 6: loss=5.572900234519603\n",
|
||||
"Epoch: 7: loss=5.564951676341915\n",
|
||||
"Epoch: 8: loss=5.558288112064614\n",
|
||||
"Epoch: 9: loss=5.552576955031129\n",
|
||||
"Epoch: 10: loss=5.547634165194347\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"output_type": "execute_result",
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"5.547634165194347"
|
||||
]
|
||||
},
|
||||
"metadata": {},
|
||||
"execution_count": 16
|
||||
}
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"id": "W8u2qXZmaHVd"
|
||||
},
|
||||
"source": [
|
||||
"## 嘗試使用 Word2Vec\n",
|
||||
"\n",
|
||||
"要使用 Word2Vec，我們來提取與詞彙表中所有單詞對應的向量：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {
|
||||
"id": "r8TatcXjkU_t"
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"vectors = torch.stack([embedder(torch.tensor(vocab[s])) for s in vocab.itos], 0)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"id": "3OcX21UOaHVd"
|
||||
},
|
||||
"source": [
|
||||
"讓我們來看看，例如，單詞 **Paris** 是如何被編碼成向量的：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {
|
||||
"colab": {
|
||||
"base_uri": "https://localhost:8080/"
|
||||
},
|
||||
"id": "bz6tAeLzieXp",
|
||||
"outputId": "5b20850e-4342-45e9-f840-cfac2b4d61d8"
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"output_type": "stream",
|
||||
"name": "stdout",
|
||||
"text": [
|
||||
"tensor([-0.0915, 2.1224, -0.0281, -0.6819, 1.1219, 0.6458, -1.3704, -1.3314,\n",
|
||||
" -1.1437, 0.4496, 0.2301, -0.3515, -0.8485, 1.0481, 0.4386, -0.8949,\n",
|
||||
" 0.5644, 1.0939, -2.5096, 3.2949, -0.2601, -0.8640, 0.1421, -0.0804,\n",
|
||||
" -0.5083, -1.0560, 0.9753, -0.5949, -1.6046, 0.5774],\n",
|
||||
" grad_fn=<EmbeddingBackward>)\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"paris_vec = embedder(torch.tensor(vocab['paris']))\n",
|
||||
"print(paris_vec)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"id": "pHTJlaeYaHVd"
|
||||
},
|
||||
"source": [
|
||||
"使用 Word2Vec 來尋找同義詞是很有趣的。以下函數將返回與給定輸入最接近的 `n` 個單詞。為了找到它們，我們計算 $|w_i - v|$ 的範數，其中 $v$ 是對應於我們輸入單詞的向量，$w_i$ 是詞彙表中第 $i$ 個單詞的編碼。我們接著對數組進行排序，並使用 `argsort` 返回相應的索引，然後取列表的前 `n` 個元素，這些元素編碼了詞彙表中最接近單詞的位置。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {
|
||||
"colab": {
|
||||
"base_uri": "https://localhost:8080/"
|
||||
},
|
||||
"id": "NlZyi-_olFar",
|
||||
"outputId": "b5dbb163-88c4-4d5a-eaf2-6751f700e98c"
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"output_type": "execute_result",
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"['microsoft', 'quoted', 'lp', 'rate', 'top']"
|
||||
]
|
||||
},
|
||||
"metadata": {},
|
||||
"execution_count": 56
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"def close_words(x, n = 5):\n",
|
||||
" vec = embedder(torch.tensor(vocab[x]))\n",
|
||||
" top5 = np.linalg.norm(vectors.detach().numpy() - vec.detach().numpy(), axis = 1).argsort()[:n]\n",
|
||||
" return [ vocab.itos[x] for x in top5 ]\n",
|
||||
"\n",
|
||||
"close_words('microsoft')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {
|
||||
"colab": {
|
||||
"base_uri": "https://localhost:8080/"
|
||||
},
|
||||
"id": "-dQq7xeAln0U",
|
||||
"outputId": "66f768c3-c248-4bfd-ce4f-c8ffc6d0dd0d"
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"output_type": "execute_result",
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"['basketball', 'lot', 'sinai', 'states', 'healthdaynews']"
|
||||
]
|
||||
},
|
||||
"metadata": {},
|
||||
"execution_count": 51
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"close_words('basketball')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {
|
||||
"colab": {
|
||||
"base_uri": "https://localhost:8080/"
|
||||
},
|
||||
"id": "fJXqK26b29sa",
|
||||
"outputId": "78f0baba-ffd0-485a-dd87-0a12bedfd7fa"
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"output_type": "execute_result",
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"['funds', 'travel', 'sydney', 'japan', 'business']"
|
||||
]
|
||||
},
|
||||
"metadata": {},
|
||||
"execution_count": 77
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"close_words('funds')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"id": "My0VeTDd3Ji8"
|
||||
},
|
||||
"source": [
|
||||
"## 重點\n",
|
||||
"\n",
|
||||
"使用像 CBoW 這樣的巧妙技術，我們可以訓練 Word2Vec 模型。你也可以嘗試訓練 skip-gram 模型，該模型是根據中心詞來預測相鄰詞，看看它的表現如何。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件使用 AI 翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。應以原文文件作為權威來源。對於關鍵資訊，建議尋求專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或誤讀概不負責。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"colab": {
|
||||
"collapsed_sections": [],
|
||||
"name": "CBoW-PyTorch.ipynb",
|
||||
"provenance": []
|
||||
},
|
||||
"interpreter": {
|
||||
"hash": "16af2a8bbb083ea23e5e41c7f5787656b2ce26968575d8763f2c4b17f9cd711f"
|
||||
},
|
||||
"kernelspec": {
|
||||
"display_name": "Python 3.8.12 ('py38')",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.8.12"
|
||||
},
|
||||
"orig_nbformat": 4,
|
||||
"gpuClass": "standard",
|
||||
"coopTranslator": {
|
||||
"original_hash": "36df28efe3fe40b6fb0a7fa48fe3ea82",
|
||||
"translation_date": "2025-08-31T10:35:52+00:00",
|
||||
"source_file": "lessons/5-NLP/15-LanguageModeling/CBoW-PyTorch.ipynb",
|
||||
"language_code": "tw"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 0
|
||||
}
|
||||
File diff suppressed because it is too large
Load Diff
|
|
@ -0,0 +1,479 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 循環神經網路\n",
|
||||
"\n",
|
||||
"在前一個模組中，我們使用了豐富的文本語義表示，並在嵌入層之上使用了一個簡單的線性分類器。這種架構的作用是捕捉句子中詞彙的聚合意義，但它並未考慮到詞彙的**順序**，因為嵌入層上的聚合操作已經移除了原始文本中的這種信息。由於這些模型無法建模詞彙的順序，因此它們無法解決更複雜或更具歧義的任務，例如文本生成或問答。\n",
|
||||
"\n",
|
||||
"為了捕捉文本序列的意義，我們需要使用另一種神經網路架構，稱為**循環神經網路**（Recurrent Neural Network，簡稱 RNN）。在 RNN 中，我們將句子逐個符號地輸入網路，網路會生成某種**狀態**，然後我們將該狀態與下一個符號一起再次輸入網路。\n",
|
||||
"\n",
|
||||
"給定輸入的序列標記 $X_0,\\dots,X_n$，RNN 會創建一個神經網路模塊的序列，並通過反向傳播對這個序列進行端到端的訓練。每個網路模塊以一對 $(X_i,S_i)$ 作為輸入，並生成 $S_{i+1}$ 作為結果。最終的狀態 $S_n$ 或輸出 $X_n$ 會進入線性分類器以生成結果。所有的網路模塊共享相同的權重，並通過一次反向傳播進行端到端的訓練。\n",
|
||||
"\n",
|
||||
"由於狀態向量 $S_0,\\dots,S_n$ 會在網路中傳遞，因此它能夠學習詞彙之間的順序依賴關係。例如，當單詞 *not* 出現在序列中的某處時，網路可以學會在狀態向量中否定某些元素，從而實現否定的效果。\n",
|
||||
"\n",
|
||||
"> 由於圖片中所有 RNN 模塊的權重是共享的，因此同一張圖片可以表示為一個帶有循環反饋迴路的模塊（如右側所示），該迴路將網路的輸出狀態反饋回輸入。\n",
|
||||
"\n",
|
||||
"接下來，我們來看看循環神經網路如何幫助我們對新聞數據集進行分類。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 1,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Loading dataset...\n",
|
||||
"Building vocab...\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"import torch\n",
|
||||
"import torchtext\n",
|
||||
"from torchnlp import *\n",
|
||||
"train_dataset, test_dataset, classes, vocab = load_dataset()\n",
|
||||
"vocab_size = len(vocab)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 簡單的 RNN 分類器\n",
|
||||
"\n",
|
||||
"在簡單的 RNN 中，每個循環單元是一個簡單的線性網路，它接收拼接後的輸入向量和狀態向量，並生成一個新的狀態向量。PyTorch 使用 `RNNCell` 類來表示這種單元，而由這些單元組成的網路則表示為 `RNN` 層。\n",
|
||||
"\n",
|
||||
"為了定義一個 RNN 分類器，我們將首先應用一個嵌入層來降低輸入詞彙的維度，然後在其上添加 RNN 層：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 2,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"class RNNClassifier(torch.nn.Module):\n",
|
||||
" def __init__(self, vocab_size, embed_dim, hidden_dim, num_class):\n",
|
||||
" super().__init__()\n",
|
||||
" self.hidden_dim = hidden_dim\n",
|
||||
" self.embedding = torch.nn.Embedding(vocab_size, embed_dim)\n",
|
||||
" self.rnn = torch.nn.RNN(embed_dim,hidden_dim,batch_first=True)\n",
|
||||
" self.fc = torch.nn.Linear(hidden_dim, num_class)\n",
|
||||
"\n",
|
||||
" def forward(self, x):\n",
|
||||
" batch_size = x.size(0)\n",
|
||||
" x = self.embedding(x)\n",
|
||||
" x,h = self.rnn(x)\n",
|
||||
" return self.fc(x.mean(dim=1))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"> **注意：** 為了簡化，我們在這裡使用未訓練的嵌入層，但如果想要更好的結果，可以使用預訓練的嵌入層，例如 Word2Vec 或 GloVe 嵌入，這在前一單元中已經描述過。為了更好地理解，您可能需要調整此程式碼以使用預訓練的嵌入。\n",
|
||||
"\n",
|
||||
"在我們的案例中，我們將使用填充的資料加載器，因此每個批次都會包含相同長度的一些填充序列。RNN 層將接收嵌入張量的序列，並產生兩個輸出：\n",
|
||||
"* $x$ 是每一步驟中 RNN 單元輸出的序列\n",
|
||||
"* $h$ 是序列最後一個元素的最終隱藏狀態\n",
|
||||
"\n",
|
||||
"接著，我們應用一個全連接的線性分類器來獲得類別數。\n",
|
||||
"\n",
|
||||
"> **注意：** RNN 的訓練相當困難，因為一旦 RNN 單元沿著序列長度展開，反向傳播所涉及的層數就會非常多。因此，我們需要選擇較小的學習率，並在更大的資料集上訓練網路以獲得良好的結果。這可能需要相當長的時間，因此建議使用 GPU。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {
|
||||
"scrolled": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"3200: acc=0.3090625\n",
|
||||
"6400: acc=0.38921875\n",
|
||||
"9600: acc=0.4590625\n",
|
||||
"12800: acc=0.511953125\n",
|
||||
"16000: acc=0.5506875\n",
|
||||
"19200: acc=0.57921875\n",
|
||||
"22400: acc=0.6070089285714285\n",
|
||||
"25600: acc=0.6304296875\n",
|
||||
"28800: acc=0.6484027777777778\n",
|
||||
"32000: acc=0.66509375\n",
|
||||
"35200: acc=0.6790056818181818\n",
|
||||
"38400: acc=0.6929166666666666\n",
|
||||
"41600: acc=0.7035817307692308\n",
|
||||
"44800: acc=0.7137276785714286\n",
|
||||
"48000: acc=0.72225\n",
|
||||
"51200: acc=0.73001953125\n",
|
||||
"54400: acc=0.7372794117647059\n",
|
||||
"57600: acc=0.7436631944444444\n",
|
||||
"60800: acc=0.7503947368421052\n",
|
||||
"64000: acc=0.75634375\n",
|
||||
"67200: acc=0.7615773809523809\n",
|
||||
"70400: acc=0.7662642045454545\n",
|
||||
"73600: acc=0.7708423913043478\n",
|
||||
"76800: acc=0.7751822916666666\n",
|
||||
"80000: acc=0.7790625\n",
|
||||
"83200: acc=0.7825\n",
|
||||
"86400: acc=0.7858564814814815\n",
|
||||
"89600: acc=0.7890513392857142\n",
|
||||
"92800: acc=0.7920474137931034\n",
|
||||
"96000: acc=0.7952708333333334\n",
|
||||
"99200: acc=0.7982258064516129\n",
|
||||
"102400: acc=0.80099609375\n",
|
||||
"105600: acc=0.8037594696969697\n",
|
||||
"108800: acc=0.8060569852941176\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=16, collate_fn=padify, shuffle=True)\n",
|
||||
"net = RNNClassifier(vocab_size,64,32,len(classes)).to(device)\n",
|
||||
"train_epoch(net,train_loader, lr=0.001)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 長短期記憶（LSTM）\n",
|
||||
"\n",
|
||||
"傳統 RNN 的主要問題之一是所謂的 **梯度消失** 問題。由於 RNN 是在一次反向傳播中端到端訓練的，因此在將誤差傳遞到網絡的第一層時會遇到困難，導致網絡無法學習遠距離的詞元之間的關係。解決這個問題的一種方法是通過使用所謂的 **門控機制** 引入 **顯式狀態管理**。這類架構中最著名的兩種是：**長短期記憶**（LSTM）和 **門控循環單元**（GRU）。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"LSTM 網絡的組織方式與 RNN 類似，但有兩個狀態會從一層傳遞到下一層：實際狀態 $c$ 和隱藏向量 $h$。在每個單元中，隱藏向量 $h_i$ 與輸入 $x_i$ 連接在一起，並通過 **門控機制** 控制狀態 $c$ 的變化。每個門控機制是一個帶有 sigmoid 激活函數（輸出範圍為 $[0,1]$）的神經網絡，可以被視為在與狀態向量相乘時的位掩碼。以下是這些門控機制（如上圖從左到右）：\n",
|
||||
"* **遺忘門** 接收隱藏向量並決定向量 $c$ 的哪些組件需要被遺忘，哪些需要保留。\n",
|
||||
"* **輸入門** 從輸入和隱藏向量中提取一些信息，並將其插入到狀態中。\n",
|
||||
"* **輸出門** 通過帶有 $\\tanh$ 激活函數的線性層轉換狀態，然後使用隱藏向量 $h_i$ 選擇其部分組件以生成新狀態 $c_{i+1}$。\n",
|
||||
"\n",
|
||||
"狀態 $c$ 的組件可以被視為一些可以開啟或關閉的標誌。例如，當我們在序列中遇到名字 *Alice* 時，我們可能會假設它指的是女性角色，並在狀態中設置一個標誌，表示句子中有一個女性名詞。當我們進一步遇到短語 *and Tom* 時，我們會設置一個標誌，表示句子中有複數名詞。因此，通過操作狀態，我們可以假設性地跟蹤句子部分的語法屬性。\n",
|
||||
"\n",
|
||||
"> **Note**: 理解 LSTM 內部結構的一個很棒的資源是 Christopher Olah 的這篇優秀文章 [Understanding LSTM Networks](https://colah.github.io/posts/2015-08-Understanding-LSTMs/)。\n",
|
||||
"\n",
|
||||
"雖然 LSTM 單元的內部結構看起來很複雜，但 PyTorch 將這些實現隱藏在 `LSTMCell` 類中，並提供 `LSTM` 對象來表示整個 LSTM 層。因此，LSTM 分類器的實現將與我們之前看到的簡單 RNN 非常相似：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"class LSTMClassifier(torch.nn.Module):\n",
|
||||
" def __init__(self, vocab_size, embed_dim, hidden_dim, num_class):\n",
|
||||
" super().__init__()\n",
|
||||
" self.hidden_dim = hidden_dim\n",
|
||||
" self.embedding = torch.nn.Embedding(vocab_size, embed_dim)\n",
|
||||
" self.embedding.weight.data = torch.randn_like(self.embedding.weight.data)-0.5\n",
|
||||
" self.rnn = torch.nn.LSTM(embed_dim,hidden_dim,batch_first=True)\n",
|
||||
" self.fc = torch.nn.Linear(hidden_dim, num_class)\n",
|
||||
"\n",
|
||||
" def forward(self, x):\n",
|
||||
" batch_size = x.size(0)\n",
|
||||
" x = self.embedding(x)\n",
|
||||
" x,(h,c) = self.rnn(x)\n",
|
||||
" return self.fc(h[-1])"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": []
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"3200: acc=0.259375\n",
|
||||
"6400: acc=0.25859375\n",
|
||||
"9600: acc=0.26177083333333334\n",
|
||||
"12800: acc=0.2784375\n",
|
||||
"16000: acc=0.313\n",
|
||||
"19200: acc=0.3528645833333333\n",
|
||||
"22400: acc=0.3965625\n",
|
||||
"25600: acc=0.4385546875\n",
|
||||
"28800: acc=0.4752777777777778\n",
|
||||
"32000: acc=0.505375\n",
|
||||
"35200: acc=0.5326704545454546\n",
|
||||
"38400: acc=0.5557552083333334\n",
|
||||
"41600: acc=0.5760817307692307\n",
|
||||
"44800: acc=0.5954910714285714\n",
|
||||
"48000: acc=0.6118333333333333\n",
|
||||
"51200: acc=0.62681640625\n",
|
||||
"54400: acc=0.6404779411764706\n",
|
||||
"57600: acc=0.6520138888888889\n",
|
||||
"60800: acc=0.662828947368421\n",
|
||||
"64000: acc=0.673546875\n",
|
||||
"67200: acc=0.6831547619047619\n",
|
||||
"70400: acc=0.6917897727272727\n",
|
||||
"73600: acc=0.6997146739130434\n",
|
||||
"76800: acc=0.707109375\n",
|
||||
"80000: acc=0.714075\n",
|
||||
"83200: acc=0.7209134615384616\n",
|
||||
"86400: acc=0.727037037037037\n",
|
||||
"89600: acc=0.7326674107142858\n",
|
||||
"92800: acc=0.7379633620689655\n",
|
||||
"96000: acc=0.7433645833333333\n",
|
||||
"99200: acc=0.7479032258064516\n",
|
||||
"102400: acc=0.752119140625\n",
|
||||
"105600: acc=0.7562405303030303\n",
|
||||
"108800: acc=0.76015625\n",
|
||||
"112000: acc=0.7641339285714286\n",
|
||||
"115200: acc=0.7677777777777778\n",
|
||||
"118400: acc=0.7711233108108108\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"(0.03487814127604167, 0.7728)"
|
||||
]
|
||||
},
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"net = LSTMClassifier(vocab_size,64,32,len(classes)).to(device)\n",
|
||||
"train_epoch(net,train_loader, lr=0.001)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 打包序列\n",
|
||||
"\n",
|
||||
"在我們的例子中，我們必須用零向量填充小批量中的所有序列。雖然這會導致一些記憶體浪費，但對於 RNN 而言，更關鍵的是為填充的輸入項目創建了額外的 RNN 單元，這些單元參與了訓練，但並未攜帶任何重要的輸入資訊。如果能僅針對實際的序列長度來訓練 RNN，效果會更好。\n",
|
||||
"\n",
|
||||
"為此，PyTorch 引入了一種特殊格式來存儲填充的序列。假設我們有一個填充過的小批量輸入，看起來像這樣：\n",
|
||||
"```\n",
|
||||
"[[1,2,3,4,5],\n",
|
||||
" [6,7,8,0,0],\n",
|
||||
" [9,0,0,0,0]]\n",
|
||||
"```\n",
|
||||
"這裡的 0 代表填充的值，而輸入序列的實際長度向量是 `[5,3,1]`。\n",
|
||||
"\n",
|
||||
"為了有效地用填充序列訓練 RNN，我們希望先用較大的小批量（`[1,6,9]`）開始訓練第一組 RNN 單元，然後結束第三個序列的處理，並繼續用較小的小批量（`[2,7]`，`[3,8]`）進行訓練，依此類推。因此，打包序列被表示為一個向量——在我們的例子中是 `[1,6,9,2,7,3,8,4,5]`，以及長度向量（`[5,3,1]`），我們可以從中輕鬆重建原始的填充小批量。\n",
|
||||
"\n",
|
||||
"要生成打包序列，我們可以使用 `torch.nn.utils.rnn.pack_padded_sequence` 函數。所有的循環層，包括 RNN、LSTM 和 GRU，都支持將打包序列作為輸入，並生成打包輸出，這些輸出可以使用 `torch.nn.utils.rnn.pad_packed_sequence` 進行解碼。\n",
|
||||
"\n",
|
||||
"為了能夠生成打包序列，我們需要將長度向量傳遞給網路，因此我們需要一個不同的函數來準備小批量：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 6,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"def pad_length(b):\n",
|
||||
" # build vectorized sequence\n",
|
||||
" v = [encode(x[1]) for x in b]\n",
|
||||
" # compute max length of a sequence in this minibatch and length sequence itself\n",
|
||||
" len_seq = list(map(len,v))\n",
|
||||
" l = max(len_seq)\n",
|
||||
" return ( # tuple of three tensors - labels, padded features, length sequence\n",
|
||||
" torch.LongTensor([t[0]-1 for t in b]),\n",
|
||||
" torch.stack([torch.nn.functional.pad(torch.tensor(t),(0,l-len(t)),mode='constant',value=0) for t in v]),\n",
|
||||
" torch.tensor(len_seq)\n",
|
||||
" )\n",
|
||||
"\n",
|
||||
"train_loader_len = torch.utils.data.DataLoader(train_dataset, batch_size=16, collate_fn=pad_length, shuffle=True)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"實際的網路結構會與上面的 `LSTMClassifier` 非常相似，但在 `forward` 傳遞時，會同時接收填充過的小批量數據和序列長度的向量。在計算嵌入層後，我們會計算打包序列，將其傳遞給 LSTM 層，然後再將結果解包回來。\n",
|
||||
"\n",
|
||||
"> **注意**：我們實際上並未使用解包後的結果 `x`，因為我們在後續的計算中使用的是隱藏層的輸出。因此，我們可以完全移除這段程式碼中的解包過程。我們之所以將其保留在這裡，是為了讓您在需要在後續計算中使用網路輸出時，能夠更輕鬆地修改這段程式碼。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"class LSTMPackClassifier(torch.nn.Module):\n",
|
||||
" def __init__(self, vocab_size, embed_dim, hidden_dim, num_class):\n",
|
||||
" super().__init__()\n",
|
||||
" self.hidden_dim = hidden_dim\n",
|
||||
" self.embedding = torch.nn.Embedding(vocab_size, embed_dim)\n",
|
||||
" self.embedding.weight.data = torch.randn_like(self.embedding.weight.data)-0.5\n",
|
||||
" self.rnn = torch.nn.LSTM(embed_dim,hidden_dim,batch_first=True)\n",
|
||||
" self.fc = torch.nn.Linear(hidden_dim, num_class)\n",
|
||||
"\n",
|
||||
" def forward(self, x, lengths):\n",
|
||||
" batch_size = x.size(0)\n",
|
||||
" x = self.embedding(x)\n",
|
||||
" pad_x = torch.nn.utils.rnn.pack_padded_sequence(x,lengths,batch_first=True,enforce_sorted=False)\n",
|
||||
" pad_x,(h,c) = self.rnn(pad_x)\n",
|
||||
" x, _ = torch.nn.utils.rnn.pad_packed_sequence(pad_x,batch_first=True)\n",
|
||||
" return self.fc(h[-1])"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": []
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 8,
|
||||
"metadata": {
|
||||
"scrolled": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"3200: acc=0.285625\n",
|
||||
"6400: acc=0.33359375\n",
|
||||
"9600: acc=0.3876041666666667\n",
|
||||
"12800: acc=0.44078125\n",
|
||||
"16000: acc=0.4825\n",
|
||||
"19200: acc=0.5235416666666667\n",
|
||||
"22400: acc=0.5559821428571429\n",
|
||||
"25600: acc=0.58609375\n",
|
||||
"28800: acc=0.6116666666666667\n",
|
||||
"32000: acc=0.63340625\n",
|
||||
"35200: acc=0.6525284090909091\n",
|
||||
"38400: acc=0.668515625\n",
|
||||
"41600: acc=0.6822596153846154\n",
|
||||
"44800: acc=0.6948214285714286\n",
|
||||
"48000: acc=0.7052708333333333\n",
|
||||
"51200: acc=0.71521484375\n",
|
||||
"54400: acc=0.7239889705882353\n",
|
||||
"57600: acc=0.7315277777777778\n",
|
||||
"60800: acc=0.7388486842105263\n",
|
||||
"64000: acc=0.74571875\n",
|
||||
"67200: acc=0.7518303571428572\n",
|
||||
"70400: acc=0.7576988636363636\n",
|
||||
"73600: acc=0.7628940217391305\n",
|
||||
"76800: acc=0.7681510416666667\n",
|
||||
"80000: acc=0.7728125\n",
|
||||
"83200: acc=0.7772235576923077\n",
|
||||
"86400: acc=0.7815393518518519\n",
|
||||
"89600: acc=0.7857700892857142\n",
|
||||
"92800: acc=0.7895043103448276\n",
|
||||
"96000: acc=0.7930520833333333\n",
|
||||
"99200: acc=0.7959072580645161\n",
|
||||
"102400: acc=0.798994140625\n",
|
||||
"105600: acc=0.802064393939394\n",
|
||||
"108800: acc=0.8051378676470589\n",
|
||||
"112000: acc=0.8077857142857143\n",
|
||||
"115200: acc=0.8104600694444445\n",
|
||||
"118400: acc=0.8128293918918919\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"(0.029785829671223958, 0.8138166666666666)"
|
||||
]
|
||||
},
|
||||
"execution_count": 8,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"net = LSTMPackClassifier(vocab_size,64,32,len(classes)).to(device)\n",
|
||||
"train_epoch_emb(net,train_loader_len, lr=0.001,use_pack_sequence=True)\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"> **注意：** 您可能已經注意到我們傳遞給訓練函數的參數 `use_pack_sequence`。目前，`pack_padded_sequence` 函數要求長度序列張量位於 CPU 設備上，因此訓練函數需要避免在訓練時將長度序列數據移動到 GPU。您可以查看 [`torchnlp.py`](../../../../../lessons/5-NLP/16-RNN/torchnlp.py) 文件中 `train_emb` 函數的實現。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 雙向與多層 RNN\n",
|
||||
"\n",
|
||||
"在我們的範例中，所有的循環神經網路都是單向運作的，從序列的開始到結束。這看起來很自然，因為它類似於我們閱讀或聆聽語音的方式。然而，在許多實際情況下，我們可以隨機存取輸入序列，因此在兩個方向上進行循環計算可能更有意義。這類網路被稱為**雙向 RNN**，可以透過在 RNN/LSTM/GRU 的建構函數中傳入參數 `bidirectional=True` 來建立。\n",
|
||||
"\n",
|
||||
"在處理雙向網路時，我們需要兩個隱藏狀態向量，每個方向各一個。PyTorch 將這些向量編碼為一個大小加倍的向量，這非常方便，因為通常我們會將最終的隱藏狀態傳遞給全連接的線性層，只需在建立該層時考慮到這個大小的增加即可。\n",
|
||||
"\n",
|
||||
"無論是單向還是雙向的循環網路，都能在序列中捕捉某些模式，並將其存儲到狀態向量中或傳遞到輸出中。與卷積網路類似，我們可以在第一層之上再構建另一個循環層，以捕捉更高層次的模式，這些模式是由第一層提取的低層次模式構成的。這引出了**多層 RNN** 的概念，它由兩層或更多的循環網路組成，前一層的輸出作為下一層的輸入。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"*圖片來源：[這篇精彩的文章](https://towardsdatascience.com/from-a-lstm-cell-to-a-multilayer-lstm-network-with-pytorch-2899eb5696f3) 作者 Fernando López*\n",
|
||||
"\n",
|
||||
"PyTorch 讓構建這類網路變得非常簡單，因為你只需在 RNN/LSTM/GRU 的建構函數中傳入參數 `num_layers`，即可自動建立多層的循環網路。這也意味著隱藏/狀態向量的大小會按比例增加，因此在處理循環層的輸出時需要考慮到這一點。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## RNNs 用於其他任務\n",
|
||||
"\n",
|
||||
"在本單元中，我們已經看到 RNNs 可以用於序列分類，但事實上，它們還能處理更多任務，例如文本生成、機器翻譯等。我們將在下一單元中探討這些任務。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件使用 AI 翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。應以原文文件作為權威來源。對於關鍵資訊，建議尋求專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或誤讀概不負責。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"interpreter": {
|
||||
"hash": "16af2a8bbb083ea23e5e41c7f5787656b2ce26968575d8763f2c4b17f9cd711f"
|
||||
},
|
||||
"kernelspec": {
|
||||
"display_name": "Python 3.8.12 ('py38')",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.8.12"
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "522ee52ae3d5ae933e283286254e9a55",
|
||||
"translation_date": "2025-08-31T10:49:00+00:00",
|
||||
"source_file": "lessons/5-NLP/16-RNN/RNNPyTorch.ipynb",
|
||||
"language_code": "tw"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 2
|
||||
}
|
||||
|
|
@ -0,0 +1,460 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 循環神經網路\n",
|
||||
"\n",
|
||||
"在上一個模組中，我們探討了文本的豐富語義表示。我們使用的架構能夠捕捉句子中單詞的聚合意義，但它並未考慮單詞的**順序**，因為嵌入後的聚合操作會將原始文本中的這些信息移除。由於這些模型無法表示單詞的順序，因此它們無法解決更複雜或更具歧義的任務，例如文本生成或問題回答。\n",
|
||||
"\n",
|
||||
"為了捕捉文本序列的意義，我們將使用一種稱為**循環神經網路**（Recurrent Neural Network，簡稱 RNN）的神經網路架構。在使用 RNN 時，我們會將句子逐個標記（token）傳遞給網路，網路會生成某種**狀態**，然後我們將該狀態與下一個標記一起再次傳遞給網路。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"給定輸入標記序列 $X_0,\\dots,X_n$，RNN 會創建一個神經網路區塊的序列，並通過反向傳播對該序列進行端到端訓練。每個網路區塊將一對 $(X_i,S_i)$ 作為輸入，並生成 $S_{i+1}$ 作為結果。最終狀態 $S_n$ 或輸出 $Y_n$ 會進入線性分類器以生成結果。所有網路區塊共享相同的權重，並通過一次反向傳播訓練完成端到端的學習。\n",
|
||||
"\n",
|
||||
"> 上圖展示了循環神經網路的展開形式（左側）以及更緊湊的循環表示形式（右側）。需要注意的是，所有 RNN 單元都具有相同的**可共享權重**。\n",
|
||||
"\n",
|
||||
"由於狀態向量 $S_0,\\dots,S_n$ 是通過網路傳遞的，RNN 能夠學習單詞之間的順序依賴性。例如，當單詞 *not* 出現在序列中的某處時，它可以學會在狀態向量中否定某些元素。\n",
|
||||
"\n",
|
||||
"在內部，每個 RNN 單元包含兩個權重矩陣：$W_H$ 和 $W_I$，以及偏置 $b$。在每個 RNN 步驟中，給定輸入 $X_i$ 和輸入狀態 $S_i$，輸出狀態的計算方式為 $S_{i+1} = f(W_H\\times S_i + W_I\\times X_i+b)$，其中 $f$ 是一個激活函數（通常是 $\\tanh$）。\n",
|
||||
"\n",
|
||||
"> 對於像文本生成（我們將在下一單元中討論）或機器翻譯這樣的問題，我們還希望在每個 RNN 步驟中獲得一些輸出值。在這種情況下，還會有另一個矩陣 $W_O$，輸出值的計算方式為 $Y_i=f(W_O\\times S_i+b_O)$。\n",
|
||||
"\n",
|
||||
"現在讓我們看看循環神經網路如何幫助我們對新聞數據集進行分類。\n",
|
||||
"\n",
|
||||
"> 在沙盒環境中，我們需要運行以下程式碼單元，以確保安裝了所需的庫並預取了數據。如果您在本地運行，則可以跳過以下程式碼單元。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 1,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import sys\n",
|
||||
"!{sys.executable} -m pip install --quiet tensorflow_datasets==4.4.0\n",
|
||||
"!cd ~ && wget -q -O - https://mslearntensorflowlp.blob.core.windows.net/data/tfds-ag-news.tgz | tar xz"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 2,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import tensorflow as tf\n",
|
||||
"from tensorflow import keras\n",
|
||||
"import tensorflow_datasets as tfds\n",
|
||||
"import numpy as np\n",
|
||||
"\n",
|
||||
"# We are going to be training pretty large models. In order not to face errors, we need\n",
|
||||
"# to set tensorflow option to grow GPU memory allocation when required\n",
|
||||
"physical_devices = tf.config.list_physical_devices('GPU') \n",
|
||||
"if len(physical_devices)>0:\n",
|
||||
" tf.config.experimental.set_memory_growth(physical_devices[0], True)\n",
|
||||
"\n",
|
||||
"ds_train, ds_test = tfds.load('ag_news_subset').values()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"nteract": {
|
||||
"transient": {
|
||||
"deleting": false
|
||||
}
|
||||
}
|
||||
},
|
||||
"source": [
|
||||
"在訓練大型模型時，GPU 記憶體分配可能會成為一個問題。我們也可能需要嘗試不同的迷你批次大小，以便數據能夠適配 GPU 記憶體，同時確保訓練速度足夠快。如果您在自己的 GPU 機器上運行此代碼，可以嘗試調整迷你批次大小來加快訓練速度。\n",
|
||||
"\n",
|
||||
"> **注意**: 某些版本的 NVidia 驅動程式已知在訓練模型後不會釋放記憶體。我們在這個筆記本中運行了幾個範例，這可能會導致在某些配置中記憶體耗盡，特別是如果您在同一個筆記本中進行自己的實驗時。如果在開始訓練模型時遇到一些奇怪的錯誤，您可能需要重新啟動筆記本的內核。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {
|
||||
"collapsed": true,
|
||||
"jupyter": {
|
||||
"outputs_hidden": false,
|
||||
"source_hidden": false
|
||||
},
|
||||
"nteract": {
|
||||
"transient": {
|
||||
"deleting": false
|
||||
}
|
||||
}
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"batch_size = 16\n",
|
||||
"embed_size = 64"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 簡單的 RNN 分類器\n",
|
||||
"\n",
|
||||
"在簡單的 RNN 中，每個循環單元都是一個簡單的線性網絡，它接收輸入向量和狀態向量，並生成新的狀態向量。在 Keras 中，可以使用 `SimpleRNN` 層來表示。\n",
|
||||
"\n",
|
||||
"雖然我們可以直接將 one-hot 編碼的標記傳遞給 RNN 層，但由於其高維度性，這並不是一個好主意。因此，我們將使用嵌入層來降低詞向量的維度，接著是 RNN 層，最後是一個 `Dense` 分類器。\n",
|
||||
"\n",
|
||||
"> **注意**：在維度不太高的情況下，例如使用字符級標記化時，直接將 one-hot 編碼的標記傳遞給 RNN 單元可能是合理的選擇。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Model: \"sequential\"\n",
|
||||
"_________________________________________________________________\n",
|
||||
"Layer (type) Output Shape Param # \n",
|
||||
"=================================================================\n",
|
||||
"text_vectorization (TextVect (None, None) 0 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"embedding (Embedding) (None, None, 64) 1280000 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"simple_rnn (SimpleRNN) (None, 16) 1296 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"dense (Dense) (None, 4) 68 \n",
|
||||
"=================================================================\n",
|
||||
"Total params: 1,281,364\n",
|
||||
"Trainable params: 1,281,364\n",
|
||||
"Non-trainable params: 0\n",
|
||||
"_________________________________________________________________\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"vocab_size = 20000\n",
|
||||
"\n",
|
||||
"vectorizer = keras.layers.experimental.preprocessing.TextVectorization(\n",
|
||||
" max_tokens=vocab_size,\n",
|
||||
" input_shape=(1,))\n",
|
||||
"\n",
|
||||
"model = keras.models.Sequential([\n",
|
||||
" vectorizer,\n",
|
||||
" keras.layers.Embedding(vocab_size, embed_size),\n",
|
||||
" keras.layers.SimpleRNN(16),\n",
|
||||
" keras.layers.Dense(4,activation='softmax')\n",
|
||||
"])\n",
|
||||
"\n",
|
||||
"model.summary()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"> **注意：** 為了簡化，我們在這裡使用未經訓練的嵌入層，但如果想要更好的結果，可以使用 Word2Vec 預訓練的嵌入層，如前一單元所述。你可以嘗試將此程式碼改寫為使用預訓練嵌入層，這會是一個很好的練習。\n",
|
||||
"\n",
|
||||
"現在讓我們來訓練 RNN。一般來說，RNN 的訓練相當困難，因為當 RNN 單元沿著序列長度展開時，涉及反向傳播的層數會非常多。因此，我們需要選擇較小的學習率，並在更大的數據集上訓練網路以獲得良好的結果。這可能需要相當長的時間，因此建議使用 GPU。\n",
|
||||
"\n",
|
||||
"為了加快速度，我們將僅使用新聞標題來訓練 RNN 模型，省略描述部分。你可以嘗試使用描述進行訓練，看看是否能讓模型成功訓練。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {
|
||||
"scrolled": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Training vectorizer\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"def extract_title(x):\n",
|
||||
" return x['title']\n",
|
||||
"\n",
|
||||
"def tupelize_title(x):\n",
|
||||
" return (extract_title(x),x['label'])\n",
|
||||
"\n",
|
||||
"print('Training vectorizer')\n",
|
||||
"vectorizer.adapt(ds_train.take(2000).map(extract_title))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 6,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"7500/7500 [==============================] - 82s 11ms/step - loss: 0.6629 - acc: 0.7623 - val_loss: 0.5559 - val_acc: 0.7995\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<tensorflow.python.keras.callbacks.History at 0x7f3e0030d350>"
|
||||
]
|
||||
},
|
||||
"execution_count": 6,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model.compile(loss='sparse_categorical_crossentropy',metrics=['acc'], optimizer='adam')\n",
|
||||
"model.fit(ds_train.map(tupelize_title).batch(batch_size),validation_data=ds_test.map(tupelize_title).batch(batch_size))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"nteract": {
|
||||
"transient": {
|
||||
"deleting": false
|
||||
}
|
||||
}
|
||||
},
|
||||
"source": [
|
||||
"> **注意** 這裡的準確性可能會較低，因為我們僅針對新聞標題進行訓練。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 重新探討變長序列\n",
|
||||
"\n",
|
||||
"請記住，`TextVectorization` 層會自動在小批量中用填充標記（pad tokens）填充變長序列。然而，這些填充標記也會參與訓練，這可能會使模型的收斂變得更加複雜。\n",
|
||||
"\n",
|
||||
"我們可以採取幾種方法來減少填充的數量。其中一種方法是根據序列長度重新排序數據集，並將所有序列按大小分組。這可以使用 `tf.data.experimental.bucket_by_sequence_length` 函數來完成（參見[文件](https://www.tensorflow.org/api_docs/python/tf/data/experimental/bucket_by_sequence_length)）。\n",
|
||||
"\n",
|
||||
"另一種方法是使用**遮罩（masking）**。在 Keras 中，一些層支持額外的輸入，用於指示哪些標記應該在訓練時被考慮。要將遮罩整合到模型中，我們可以選擇加入一個單獨的 `Masking` 層（[文件](https://keras.io/api/layers/core_layers/masking/)），或者在 `Embedding` 層中指定參數 `mask_zero=True`。\n",
|
||||
"\n",
|
||||
"> **Note**: 完成整個數據集的一個訓練週期大約需要 5 分鐘。如果你失去耐心，可以隨時中斷訓練。你還可以通過在 `ds_train` 和 `ds_test` 數據集後添加 `.take(...)` 子句來限制用於訓練的數據量。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"7500/7500 [==============================] - 371s 49ms/step - loss: 0.5401 - acc: 0.8079 - val_loss: 0.3780 - val_acc: 0.8822\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<tensorflow.python.keras.callbacks.History at 0x7f3dec118850>"
|
||||
]
|
||||
},
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"def extract_text(x):\n",
|
||||
" return x['title']+' '+x['description']\n",
|
||||
"\n",
|
||||
"def tupelize(x):\n",
|
||||
" return (extract_text(x),x['label'])\n",
|
||||
"\n",
|
||||
"model = keras.models.Sequential([\n",
|
||||
" vectorizer,\n",
|
||||
" keras.layers.Embedding(vocab_size,embed_size,mask_zero=True),\n",
|
||||
" keras.layers.SimpleRNN(16),\n",
|
||||
" keras.layers.Dense(4,activation='softmax')\n",
|
||||
"])\n",
|
||||
"\n",
|
||||
"model.compile(loss='sparse_categorical_crossentropy',metrics=['acc'], optimizer='adam')\n",
|
||||
"model.fit(ds_train.map(tupelize).batch(batch_size),validation_data=ds_test.map(tupelize).batch(batch_size))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在我們使用遮罩技術，可以在整個標題和描述的數據集上訓練模型。\n",
|
||||
"\n",
|
||||
"> **注意**：你是否注意到我們一直在使用基於新聞標題訓練的向量化工具，而不是整篇文章的內容？這可能會導致某些詞元被忽略，因此重新訓練向量化工具會更好。不過，這可能只會帶來非常小的影響，所以為了簡化流程，我們將繼續使用之前預訓練的向量化工具。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## LSTM: 長短期記憶\n",
|
||||
"\n",
|
||||
"RNN 的主要問題之一是**梯度消失**。RNN 可能會非常長，並且在反向傳播過程中，可能很難將梯度傳遞回網路的第一層。當這種情況發生時，網路無法學習遠距離的詞元之間的關係。為了避免這個問題，可以通過使用**門控機制**引入**顯式狀態管理**。最常見的兩種引入門控機制的架構是**長短期記憶**（LSTM）和**門控循環單元**（GRU）。我們在這裡將介紹 LSTM。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"LSTM 網路的組織方式與 RNN 類似，但有兩個狀態會從一層傳遞到下一層：實際狀態 $c$ 和隱藏向量 $h$。在每個單元中，隱藏向量 $h_{t-1}$ 與輸入 $x_t$ 結合，並共同控制狀態 $c_t$ 和輸出 $h_{t}$ 的變化，這是通過**門控機制**實現的。每個門都有 sigmoid 激活函數（輸出範圍為 $[0,1]$），可以將其視為與狀態向量相乘的位掩碼。LSTM 包含以下門控機制（如上圖從左到右）：\n",
|
||||
"* **遺忘門**：決定向量 $c_{t-1}$ 的哪些部分需要遺忘，哪些需要保留。\n",
|
||||
"* **輸入門**：決定來自輸入向量和前一隱藏向量的信息有多少應該被整合到狀態向量中。\n",
|
||||
"* **輸出門**：接收新的狀態向量，並決定其哪些部分將用於生成新的隱藏向量 $h_t$。\n",
|
||||
"\n",
|
||||
"狀態 $c$ 的組成部分可以被視為可以開啟或關閉的標誌。例如，當我們在序列中遇到名字 *Alice* 時，我們猜測它指的是一位女性，並在狀態中設置一個標誌，表示句子中有一個女性名詞。當我們進一步遇到單詞 *and Tom* 時，我們會設置一個標誌，表示句子中有一個複數名詞。因此，通過操作狀態，我們可以追蹤句子的語法屬性。\n",
|
||||
"\n",
|
||||
"> **Note**: 這裡有一個很棒的資源可以幫助理解 LSTM 的內部結構：[Understanding LSTM Networks](https://colah.github.io/posts/2015-08-Understanding-LSTMs/) by Christopher Olah。\n",
|
||||
"\n",
|
||||
"雖然 LSTM 單元的內部結構看起來可能很複雜，但 Keras 將這些實現隱藏在 `LSTM` 層中，因此在上面的範例中，我們只需要替換循環層即可：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 8,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"15000/15000 [==============================] - 188s 13ms/step - loss: 0.5692 - acc: 0.7916 - val_loss: 0.3441 - val_acc: 0.8870\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<tensorflow.python.keras.callbacks.History at 0x7f3d6af5c350>"
|
||||
]
|
||||
},
|
||||
"execution_count": 8,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model = keras.models.Sequential([\n",
|
||||
" vectorizer,\n",
|
||||
" keras.layers.Embedding(vocab_size, embed_size),\n",
|
||||
" keras.layers.LSTM(8),\n",
|
||||
" keras.layers.Dense(4,activation='softmax')\n",
|
||||
"])\n",
|
||||
"\n",
|
||||
"model.compile(loss='sparse_categorical_crossentropy',metrics=['acc'], optimizer='adam')\n",
|
||||
"model.fit(ds_train.map(tupelize).batch(8),validation_data=ds_test.map(tupelize).batch(8))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": []
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 雙向與多層 RNN\n",
|
||||
"\n",
|
||||
"在我們之前的範例中，循環神經網路都是從序列的開頭運算到結尾。這對我們來說很自然，因為它遵循了我們閱讀或聆聽語音的方向。然而，對於需要隨機存取輸入序列的情境，讓循環運算在兩個方向上進行會更合理。允許在兩個方向上進行運算的 RNN 被稱為 **雙向** RNN，可以透過將循環層包裹在特殊的 `Bidirectional` 層中來建立。\n",
|
||||
"\n",
|
||||
"> **Note**: `Bidirectional` 層會在其內部建立該層的兩個副本，並將其中一個副本的 `go_backwards` 屬性設置為 `True`，使其沿著序列的相反方向運算。\n",
|
||||
"\n",
|
||||
"無論是單向還是雙向的循環神經網路，都能捕捉序列中的模式，並將其存儲到狀態向量中或作為輸出返回。與卷積神經網路類似，我們可以在第一層之後再建立另一個循環層，以捕捉更高層次的模式，這些模式是由第一層提取的低層次模式構建而成的。這引出了 **多層 RNN** 的概念，它由兩層或更多層循環神經網路組成，其中前一層的輸出作為下一層的輸入。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"*圖片來源：[這篇精彩文章](https://towardsdatascience.com/from-a-lstm-cell-to-a-multilayer-lstm-network-with-pytorch-2899eb5696f3) 作者 Fernando López。*\n",
|
||||
"\n",
|
||||
"Keras 讓構建這些網路變得非常簡單，因為你只需要在模型中添加更多的循環層。對於除了最後一層以外的所有層，我們需要指定 `return_sequences=True` 參數，因為我們需要該層返回所有中間狀態，而不僅僅是循環運算的最終狀態。\n",
|
||||
"\n",
|
||||
"現在我們來為分類問題構建一個雙層雙向 LSTM。\n",
|
||||
"\n",
|
||||
"> **Note** 這段程式碼執行時間可能會比較長，但它提供了我們目前看到的最高準確率。所以也許值得等待並查看結果。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 9,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"5044/7500 [===================>..........] - ETA: 2:33 - loss: 0.3709 - acc: 0.8706\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\b\r5045/7500 [===================>..........] - ETA: 2:33 - loss: 0.3709 - acc: 0.8706"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model = keras.models.Sequential([\n",
|
||||
" vectorizer,\n",
|
||||
" keras.layers.Embedding(vocab_size, 128, mask_zero=True),\n",
|
||||
" keras.layers.Bidirectional(keras.layers.LSTM(64,return_sequences=True)),\n",
|
||||
" keras.layers.Bidirectional(keras.layers.LSTM(64)), \n",
|
||||
" keras.layers.Dense(4,activation='softmax')\n",
|
||||
"])\n",
|
||||
"\n",
|
||||
"model.compile(loss='sparse_categorical_crossentropy',metrics=['acc'], optimizer='adam')\n",
|
||||
"model.fit(ds_train.map(tupelize).batch(batch_size),\n",
|
||||
" validation_data=ds_test.map(tupelize).batch(batch_size))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## RNN 用於其他任務\n",
|
||||
"\n",
|
||||
"到目前為止，我們專注於使用 RNN 來對文本序列進行分類。但它們還可以處理更多任務，例如文本生成和機器翻譯——我們將在下一單元中探討這些任務。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件使用 AI 翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。儘管我們努力確保翻譯的準確性，但請注意，自動翻譯可能包含錯誤或不準確之處。應以原始語言的文件作為權威來源。對於關鍵資訊，建議尋求專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或誤讀概不負責。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"kernel_info": {
|
||||
"name": "conda-env-py37_tensorflow-py"
|
||||
},
|
||||
"kernelspec": {
|
||||
"display_name": "py37_tensorflow",
|
||||
"language": "python",
|
||||
"name": "conda-env-py37_tensorflow-py"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.7.9"
|
||||
},
|
||||
"nteract": {
|
||||
"version": "nteract-front-end@1.0.0"
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "81351e61f619b432ff51010a4f993194",
|
||||
"translation_date": "2025-08-31T10:46:36+00:00",
|
||||
"source_file": "lessons/5-NLP/16-RNN/RNNTF.ipynb",
|
||||
"language_code": "tw"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 4
|
||||
}
|
||||
|
|
@ -0,0 +1,414 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 生成式網絡\n",
|
||||
"\n",
|
||||
"循環神經網絡（Recurrent Neural Networks, RNNs）及其門控單元變體，例如長短期記憶單元（Long Short Term Memory Cells, LSTMs）和門控循環單元（Gated Recurrent Units, GRUs），提供了一種語言建模的機制，也就是說，它們可以學習單詞的排列順序，並對序列中的下一個單詞進行預測。這使得我們可以使用 RNNs 來完成**生成任務**，例如普通文本生成、機器翻譯，甚至是圖像描述生成。\n",
|
||||
"\n",
|
||||
"在我們上一單元討論的 RNN 架構中，每個 RNN 單元會生成下一個隱藏狀態作為輸出。然而，我們也可以為每個循環單元添加另一個輸出，這樣就可以輸出一個**序列**（其長度等於原始序列的長度）。此外，我們還可以使用不在每一步接受輸入的 RNN 單元，而僅僅接受一個初始狀態向量，然後生成一個輸出序列。\n",
|
||||
"\n",
|
||||
"在這個筆記本中，我們將專注於幫助我們生成文本的簡單生成模型。為了簡化，我們將構建一個**字符級網絡**，逐字母生成文本。在訓練過程中，我們需要採用一些文本語料庫，並將其拆分為字母序列。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 1,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Loading dataset...\n",
|
||||
"Building vocab...\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"import torch\n",
|
||||
"import torchtext\n",
|
||||
"import numpy as np\n",
|
||||
"from torchnlp import *\n",
|
||||
"train_dataset,test_dataset,classes,vocab = load_dataset()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 建立字符詞彙表\n",
|
||||
"\n",
|
||||
"為了構建字符級生成網絡，我們需要將文本拆分為單個字符，而不是單詞。這可以通過定義一個不同的分詞器來完成：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 2,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Vocabulary size = 82\n",
|
||||
"Encoding of 'a' is 1\n",
|
||||
"Character with code 13 is c\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"def char_tokenizer(words):\n",
|
||||
" return list(words) #[word for word in words]\n",
|
||||
"\n",
|
||||
"counter = collections.Counter()\n",
|
||||
"for (label, line) in train_dataset:\n",
|
||||
" counter.update(char_tokenizer(line))\n",
|
||||
"vocab = torchtext.vocab.vocab(counter)\n",
|
||||
"\n",
|
||||
"vocab_size = len(vocab)\n",
|
||||
"print(f\"Vocabulary size = {vocab_size}\")\n",
|
||||
"print(f\"Encoding of 'a' is {vocab.get_stoi()['a']}\")\n",
|
||||
"print(f\"Character with code 13 is {vocab.get_itos()[13]}\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"讓我們看看如何對我們數據集中的文本進行編碼的示例：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"tensor([ 0, 1, 2, 2, 3, 4, 5, 6, 3, 7, 8, 1, 9, 10, 3, 11, 2, 1,\n",
|
||||
" 12, 3, 7, 1, 13, 14, 3, 15, 16, 5, 17, 3, 5, 18, 8, 3, 7, 2,\n",
|
||||
" 1, 13, 14, 3, 19, 20, 8, 21, 5, 8, 9, 10, 22, 3, 20, 8, 21, 5,\n",
|
||||
" 8, 9, 10, 3, 23, 3, 4, 18, 17, 9, 5, 23, 10, 8, 2, 2, 8, 9,\n",
|
||||
" 10, 24, 3, 0, 1, 2, 2, 3, 4, 5, 9, 8, 8, 5, 25, 10, 3, 26,\n",
|
||||
" 12, 27, 16, 26, 2, 27, 16, 28, 29, 30, 1, 16, 26, 3, 17, 31, 3, 21,\n",
|
||||
" 2, 5, 9, 1, 23, 13, 32, 16, 27, 13, 10, 24, 3, 1, 9, 8, 3, 10,\n",
|
||||
" 8, 8, 27, 16, 28, 3, 28, 9, 8, 8, 16, 3, 1, 28, 1, 27, 16, 6])"
|
||||
]
|
||||
},
|
||||
"execution_count": 3,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"def enc(x):\n",
|
||||
" return torch.LongTensor(encode(x,voc=vocab,tokenizer=char_tokenizer))\n",
|
||||
"\n",
|
||||
"enc(train_dataset[0][1])"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 訓練生成式 RNN\n",
|
||||
"\n",
|
||||
"我們訓練 RNN 生成文本的方式如下。在每一步中，我們會取一段長度為 `nchars` 的字符序列，並要求網絡為每個輸入字符生成下一個輸出字符：\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"根據實際情況，我們可能還需要加入一些特殊字符，例如 *序列結束* `<eos>`。在我們的案例中，我們只希望訓練網絡進行無限文本生成，因此我們將每個序列的大小固定為 `nchars` 個標記。因此，每個訓練樣本將由 `nchars` 個輸入和 `nchars` 個輸出組成（輸出序列是輸入序列向左移動一個符號）。小批量訓練將由多個這樣的序列組成。\n",
|
||||
"\n",
|
||||
"我們生成小批量的方式是取每段長度為 `l` 的新聞文本，並從中生成所有可能的輸入-輸出組合（將有 `l-nchars` 個這樣的組合）。這些組合將形成一個小批量，而每次訓練步驟的小批量大小會有所不同。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"(tensor([[ 0, 1, 2, ..., 28, 29, 30],\n",
|
||||
" [ 1, 2, 2, ..., 29, 30, 1],\n",
|
||||
" [ 2, 2, 3, ..., 30, 1, 16],\n",
|
||||
" ...,\n",
|
||||
" [20, 8, 21, ..., 1, 28, 1],\n",
|
||||
" [ 8, 21, 5, ..., 28, 1, 27],\n",
|
||||
" [21, 5, 8, ..., 1, 27, 16]]),\n",
|
||||
" tensor([[ 1, 2, 2, ..., 29, 30, 1],\n",
|
||||
" [ 2, 2, 3, ..., 30, 1, 16],\n",
|
||||
" [ 2, 3, 4, ..., 1, 16, 26],\n",
|
||||
" ...,\n",
|
||||
" [ 8, 21, 5, ..., 28, 1, 27],\n",
|
||||
" [21, 5, 8, ..., 1, 27, 16],\n",
|
||||
" [ 5, 8, 9, ..., 27, 16, 6]]))"
|
||||
]
|
||||
},
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"nchars = 100\n",
|
||||
"\n",
|
||||
"def get_batch(s,nchars=nchars):\n",
|
||||
" ins = torch.zeros(len(s)-nchars,nchars,dtype=torch.long,device=device)\n",
|
||||
" outs = torch.zeros(len(s)-nchars,nchars,dtype=torch.long,device=device)\n",
|
||||
" for i in range(len(s)-nchars):\n",
|
||||
" ins[i] = enc(s[i:i+nchars])\n",
|
||||
" outs[i] = enc(s[i+1:i+nchars+1])\n",
|
||||
" return ins,outs\n",
|
||||
"\n",
|
||||
"get_batch(train_dataset[0][1])"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在讓我們定義生成器網絡。它可以基於我們在上一單元中討論過的任何循環單元（簡單、LSTM 或 GRU）。在我們的例子中，我們將使用 LSTM。\n",
|
||||
"\n",
|
||||
"由於網絡以字符作為輸入，且詞彙量相對較小，因此我們不需要嵌入層，直接使用獨熱編碼的輸入即可傳遞到 LSTM 單元。然而，因為我們以字符編號作為輸入，所以在傳遞到 LSTM 之前需要先進行獨熱編碼。這可以在 `forward` 傳遞過程中通過調用 `one_hot` 函數來完成。輸出編碼器將是一個線性層，用於將隱藏狀態轉換為獨熱編碼的輸出。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"class LSTMGenerator(torch.nn.Module):\n",
|
||||
" def __init__(self, vocab_size, hidden_dim):\n",
|
||||
" super().__init__()\n",
|
||||
" self.rnn = torch.nn.LSTM(vocab_size,hidden_dim,batch_first=True)\n",
|
||||
" self.fc = torch.nn.Linear(hidden_dim, vocab_size)\n",
|
||||
"\n",
|
||||
" def forward(self, x, s=None):\n",
|
||||
" x = torch.nn.functional.one_hot(x,vocab_size).to(torch.float32)\n",
|
||||
" x,s = self.rnn(x,s)\n",
|
||||
" return self.fc(x),s"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"在訓練過程中，我們希望能夠抽樣生成的文本。為此，我們將定義 `generate` 函數，該函數會生成長度為 `size` 的輸出字串，並以初始字串 `start` 作為起點。\n",
|
||||
"\n",
|
||||
"其運作方式如下：首先，我們將整個初始字串通過網絡，並獲得輸出狀態 `s` 和下一個預測字元 `out`。由於 `out` 是獨熱編碼（one-hot encoded），我們使用 `argmax` 來獲取字元在詞彙表中的索引 `nc`，然後使用 `itos` 找出實際字元，並將其附加到結果字元列表 `chars` 中。這個生成一個字元的過程會重複執行 `size` 次，以生成所需數量的字元。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 8,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"def generate(net,size=100,start='today '):\n",
|
||||
" chars = list(start)\n",
|
||||
" out, s = net(enc(chars).view(1,-1).to(device))\n",
|
||||
" for i in range(size):\n",
|
||||
" nc = torch.argmax(out[0][-1])\n",
|
||||
" chars.append(vocab.get_itos()[nc])\n",
|
||||
" out, s = net(nc.view(1,-1),s)\n",
|
||||
" return ''.join(chars)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在開始訓練吧！訓練迴圈幾乎與我們之前的例子相同，但這次我們每隔 1000 個 epoch 就會輸出一段生成的文字，而不是準確率。\n",
|
||||
"\n",
|
||||
"需要特別注意的是我們計算損失的方式。我們需要基於單熱編碼的輸出 `out` 和期望的文字輸出 `text_out`（即字符索引的列表）來計算損失。幸運的是，`cross_entropy` 函數的第一個參數是未正規化的網路輸出，第二個參數是類別編號，這正好符合我們的需求。該函數還會自動對小批量的大小進行平均。\n",
|
||||
"\n",
|
||||
"我們還通過 `samples_to_train` 限制了訓練樣本的數量，以免等待時間過長。我們鼓勵你進行實驗，嘗試更長時間的訓練，可能是多個 epoch（在這種情況下，你需要在這段程式碼外再建立一個迴圈）。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 9,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Current loss = 4.398899078369141\n",
|
||||
"today sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr sr s\n",
|
||||
"Current loss = 2.161320447921753\n",
|
||||
"today and to the tor to to the tor to to the tor to to the tor to to the tor to to the tor to to the tor t\n",
|
||||
"Current loss = 1.6722588539123535\n",
|
||||
"today and the court to the could to the could to the could to the could to the could to the could to the c\n",
|
||||
"Current loss = 2.423795223236084\n",
|
||||
"today and a second to the conternation of the conternation of the conternation of the conternation of the \n",
|
||||
"Current loss = 1.702607274055481\n",
|
||||
"today and the company to the company to the company to the company to the company to the company to the co\n",
|
||||
"Current loss = 1.692358136177063\n",
|
||||
"today and the company to the company to the company to the company to the company to the company to the co\n",
|
||||
"Current loss = 1.9722288846969604\n",
|
||||
"today and the control the control the control the control the control the control the control the control \n",
|
||||
"Current loss = 1.8705692291259766\n",
|
||||
"today and the second to the second to the second to the second to the second to the second to the second t\n",
|
||||
"Current loss = 1.7626899480819702\n",
|
||||
"today and a security and a security and a security and a security and a security and a security and a secu\n",
|
||||
"Current loss = 1.5574463605880737\n",
|
||||
"today and the company and the company and the company and the company and the company and the company and \n",
|
||||
"Current loss = 1.5620026588439941\n",
|
||||
"today and the be that the be the be that the be the be that the be the be that the be the be that the be t\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"net = LSTMGenerator(vocab_size,64).to(device)\n",
|
||||
"\n",
|
||||
"samples_to_train = 10000\n",
|
||||
"optimizer = torch.optim.Adam(net.parameters(),0.01)\n",
|
||||
"loss_fn = torch.nn.CrossEntropyLoss()\n",
|
||||
"net.train()\n",
|
||||
"for i,x in enumerate(train_dataset):\n",
|
||||
" # x[0] is class label, x[1] is text\n",
|
||||
" if len(x[1])-nchars<10:\n",
|
||||
" continue\n",
|
||||
" samples_to_train-=1\n",
|
||||
" if not samples_to_train: break\n",
|
||||
" text_in, text_out = get_batch(x[1])\n",
|
||||
" optimizer.zero_grad()\n",
|
||||
" out,s = net(text_in)\n",
|
||||
" loss = torch.nn.functional.cross_entropy(out.view(-1,vocab_size),text_out.flatten()) #cross_entropy(out,labels)\n",
|
||||
" loss.backward()\n",
|
||||
" optimizer.step()\n",
|
||||
" if i%1000==0:\n",
|
||||
" print(f\"Current loss = {loss.item()}\")\n",
|
||||
" print(generate(net))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"這個範例已經生成了一些相當不錯的文本，但仍有幾個方面可以進一步改進：\n",
|
||||
"\n",
|
||||
"* **更好的小批次生成**。我們在準備訓練數據時，是從一個樣本中生成一個小批次。這並不理想，因為小批次的大小各不相同，有些甚至無法生成，因為文本小於 `nchars`。此外，過小的小批次無法充分利用 GPU。更明智的做法是從所有樣本中提取一大段文本，然後生成所有的輸入-輸出對，將它們打亂，並生成大小相等的小批次。\n",
|
||||
"\n",
|
||||
"* **多層 LSTM**。嘗試使用 2 或 3 層的 LSTM 單元是有意義的。如我們在前一單元提到的，每層 LSTM 都會從文本中提取特定的模式，而在字元級生成器的情況下，我們可以預期較低層的 LSTM 負責提取音節，而較高層則負責提取單詞及單詞組合。這可以通過向 LSTM 構造函數傳遞層數參數來簡單實現。\n",
|
||||
"\n",
|
||||
"* 你也可以嘗試使用 **GRU 單元**，看看哪種表現更好，以及嘗試 **不同的隱藏層大小**。過大的隱藏層可能導致過擬合（例如，網絡會學習到精確的文本），而過小的隱藏層可能無法生成良好的結果。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 軟性文本生成與溫度\n",
|
||||
"\n",
|
||||
"在之前 `generate` 的定義中，我們總是選擇機率最高的字元作為生成文本的下一個字元。這導致文本經常在相同的字元序列之間反覆循環，如以下範例所示：\n",
|
||||
"```\n",
|
||||
"today of the second the company and a second the company ...\n",
|
||||
"```\n",
|
||||
"\n",
|
||||
"然而，如果我們查看下一個字元的機率分佈，可能會發現幾個最高機率之間的差距並不大，例如一個字元的機率可能是 0.2，另一個則是 0.19，等等。例如，當尋找序列 '*play*' 的下一個字元時，下一個字元可能同樣是空格或 **e**（如單字 *player* 中）。\n",
|
||||
"\n",
|
||||
"這讓我們得出結論，選擇機率最高的字元並不總是「公平」的，因為選擇第二高的字元仍然可能生成有意義的文本。更明智的做法是根據網絡輸出的機率分佈來**抽樣**字元。\n",
|
||||
"\n",
|
||||
"這種抽樣可以使用 `multinomial` 函數來完成，該函數實現了所謂的**多項分佈**。以下定義了一個實現此**軟性**文本生成的函數：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 10,
|
||||
"metadata": {
|
||||
"scrolled": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"--- Temperature = 0.3\n",
|
||||
"Today and a company and complete an all the land the restrational the as a security and has provers the pay to and a report and the computer in the stand has filities and working the law the stations for a company and with the company and the final the first company and refight of the state and and workin\n",
|
||||
"\n",
|
||||
"--- Temperature = 0.8\n",
|
||||
"Today he oniis its first to Aus bomblaties the marmation a to manan boogot that pirate assaid a relaid their that goverfin the the Cappets Ecrotional Assonia Cition targets it annight the w scyments Blamity #39;s TVeer Diercheg Reserals fran envyuil that of ster said access what succers of Dour-provelith\n",
|
||||
"\n",
|
||||
"--- Temperature = 1.0\n",
|
||||
"Today holy they a 11 will meda a toket subsuaties, engins for Chanos, they's has stainger past to opening orital his thempting new Nattona was al innerforder advan-than #36;s night year his religuled talitatian what the but with Wednesday to Justment will wemen of Mark CCC Camp as Timed Nae wome a leaders\n",
|
||||
"\n",
|
||||
"--- Temperature = 1.3\n",
|
||||
"Today gpone 2.5 fech atcusion poor cocles toparsdorM.cht Line Pamage put 43 his calt lowed to the book, that has authh-the silia rruch ailing to'ory andhes beutirsimi- Aefffive heading offil an auf eacklets is charged evis, Gunymy oy) Mony has it after-sloythyor loveId out filme, the Natabl -Najuntaxiggs \n",
|
||||
"\n",
|
||||
"--- Temperature = 1.8\n",
|
||||
"Today plary, P.slan chly\\401 mardregationly #39;t 8.1Mide) closes ,filtcon alfly playin roven!\\grea.-QFBEP: Iss onfarchQ/itilia CCf Zivesigntwasta orce.-Peul-aw.uicrin of fuglinfsut aftaningwo, MIEX awayew Aice Woiduar Corvagiugge oppo esig ThusBratourid canthly-RyI.co lagitems\\eexciaishes.conBabntusmor I\n",
|
||||
"\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"def generate_soft(net,size=100,start='today ',temperature=1.0):\n",
|
||||
" chars = list(start)\n",
|
||||
" out, s = net(enc(chars).view(1,-1).to(device))\n",
|
||||
" for i in range(size):\n",
|
||||
" #nc = torch.argmax(out[0][-1])\n",
|
||||
" out_dist = out[0][-1].div(temperature).exp()\n",
|
||||
" nc = torch.multinomial(out_dist,1)[0]\n",
|
||||
" chars.append(vocab.get_itos()[nc])\n",
|
||||
" out, s = net(nc.view(1,-1),s)\n",
|
||||
" return ''.join(chars)\n",
|
||||
" \n",
|
||||
"for i in [0.3,0.8,1.0,1.3,1.8]:\n",
|
||||
" print(f\"--- Temperature = {i}\\n{generate_soft(net,size=300,start='Today ',temperature=i)}\\n\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們引入了一個名為 **temperature** 的參數，用於指示我們應該多大程度地堅持最高概率。如果 temperature 為 1.0，我們進行公平的多項式抽樣，而當 temperature 趨於無窮大時，所有概率變得相等，我們隨機選擇下一個字符。在下面的例子中，我們可以觀察到，當我們將 temperature 增加得過高時，文本變得毫無意義，而當它接近 0 時，則類似於「循環」的硬生成文本。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件使用 AI 翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。應以原文文件作為權威來源。對於關鍵資訊，建議尋求專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或誤讀概不負責。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"interpreter": {
|
||||
"hash": "16af2a8bbb083ea23e5e41c7f5787656b2ce26968575d8763f2c4b17f9cd711f"
|
||||
},
|
||||
"kernelspec": {
|
||||
"display_name": "Python 3.8.12 ('py38')",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.8.12"
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "7673cd150d96c74c6d6011460094efb4",
|
||||
"translation_date": "2025-08-31T10:34:02+00:00",
|
||||
"source_file": "lessons/5-NLP/17-GenerativeNetworks/GenerativePyTorch.ipynb",
|
||||
"language_code": "tw"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 4
|
||||
}
|
||||
|
|
@ -0,0 +1,497 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 生成式網絡\n",
|
||||
"\n",
|
||||
"循環神經網絡（Recurrent Neural Networks, RNNs）及其門控單元變體，例如長短期記憶單元（Long Short Term Memory Cells, LSTMs）和門控循環單元（Gated Recurrent Units, GRUs），提供了一種語言建模的機制，也就是說，它們可以學習單詞的排列順序，並對序列中的下一個單詞進行預測。這使得我們可以使用 RNNs 來完成**生成任務**，例如普通文本生成、機器翻譯，甚至是圖像描述生成。\n",
|
||||
"\n",
|
||||
"在我們上一單元討論的 RNN 架構中，每個 RNN 單元會生成下一個隱藏狀態作為輸出。然而，我們也可以為每個循環單元添加另一個輸出，這樣就可以輸出一個**序列**（其長度等於原始序列的長度）。此外，我們還可以使用不在每一步接受輸入的 RNN 單元，而僅僅接受一個初始狀態向量，然後生成一個輸出序列。\n",
|
||||
"\n",
|
||||
"在這個筆記本中，我們將專注於幫助我們生成文本的簡單生成模型。為了簡化起見，我們將構建一個**字符級網絡**，逐字母生成文本。在訓練過程中，我們需要使用一些文本語料庫，並將其拆分為字母序列。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 1,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import tensorflow as tf\n",
|
||||
"from tensorflow import keras\n",
|
||||
"import tensorflow_datasets as tfds\n",
|
||||
"import numpy as np\n",
|
||||
"\n",
|
||||
"ds_train, ds_test = tfds.load('ag_news_subset').values()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 建立字元詞彙表\n",
|
||||
"\n",
|
||||
"為了建立字元級別的生成網絡，我們需要將文本拆分為單個字元，而不是單詞。我們之前使用的 `TextVectorization` 層無法做到這一點，因此我們有以下兩個選擇：\n",
|
||||
"\n",
|
||||
"* 手動加載文本並自行進行分詞，如 [這個官方 Keras 範例](https://keras.io/examples/generative/lstm_character_level_text_generation/) 中所示\n",
|
||||
"* 使用 `Tokenizer` 類進行字元級別的分詞。\n",
|
||||
"\n",
|
||||
"我們將選擇第二種方法。`Tokenizer` 也可以用於將文本分詞為單詞，因此可以很輕鬆地從字元級分詞切換到單詞級分詞。\n",
|
||||
"\n",
|
||||
"要進行字元級分詞，我們需要傳遞參數 `char_level=True`：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 2,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"def extract_text(x):\n",
|
||||
" return x['title']+' '+x['description']\n",
|
||||
"\n",
|
||||
"def tupelize(x):\n",
|
||||
" return (extract_text(x),x['label'])\n",
|
||||
"\n",
|
||||
"tokenizer = keras.preprocessing.text.Tokenizer(char_level=True,lower=False)\n",
|
||||
"tokenizer.fit_on_texts([x['title'].numpy().decode('utf-8') for x in ds_train])"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們還希望使用一個特殊的標記來表示**序列結束**，我們將其稱為`<eos>`。讓我們手動將其添加到詞彙表中：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"eos_token = len(tokenizer.word_index)+1\n",
|
||||
"tokenizer.word_index['<eos>'] = eos_token\n",
|
||||
"\n",
|
||||
"vocab_size = eos_token + 1"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在，要將文本編碼為數字序列，我們可以使用：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"[[48, 2, 10, 10, 5, 44, 1, 25, 5, 8, 10, 13, 78]]"
|
||||
]
|
||||
},
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"tokenizer.texts_to_sequences(['Hello, world!'])"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 訓練生成式 RNN 來生成標題\n",
|
||||
"\n",
|
||||
"我們將以以下方式訓練 RNN 來生成新聞標題。在每一步中，我們會選取一個標題，將其輸入到 RNN 中，並且對於每個輸入的字元，我們會要求網路生成下一個輸出的字元：\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"對於序列中的最後一個字元，我們會要求網路生成 `<eos>` 標記。\n",
|
||||
"\n",
|
||||
"我們在此使用的生成式 RNN 的主要不同之處在於，我們會從 RNN 的每一步輸出中提取結果，而不僅僅是從最後一個單元提取。這可以通過向 RNN 單元指定 `return_sequences` 參數來實現。\n",
|
||||
"\n",
|
||||
"因此，在訓練過程中，網路的輸入將是一段編碼字元的序列，而輸出將是相同長度的序列，但向右偏移一個元素並以 `<eos>` 結束。小批次將由多個這樣的序列組成，我們需要使用**填充**來對齊所有序列。\n",
|
||||
"\n",
|
||||
"接下來，我們來建立一些函數，用於轉換數據集。由於我們希望在小批次層級進行序列填充，我們會先通過調用 `.batch()` 將數據集分批，然後使用 `map` 來進行轉換。因此，轉換函數將以整個小批次作為參數：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"def title_batch(x):\n",
|
||||
" x = [t.numpy().decode('utf-8') for t in x]\n",
|
||||
" z = tokenizer.texts_to_sequences(x)\n",
|
||||
" z = tf.keras.preprocessing.sequence.pad_sequences(z)\n",
|
||||
" return tf.one_hot(z,vocab_size), tf.one_hot(tf.concat([z[:,1:],tf.constant(eos_token,shape=(len(z),1))],axis=1),vocab_size)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"一些我們在這裡執行的重要步驟：\n",
|
||||
"* 我們首先從字串張量中提取實際的文字\n",
|
||||
"* `text_to_sequences` 將字串列表轉換為整數張量列表\n",
|
||||
"* `pad_sequences` 將這些張量填充到它們的最大長度\n",
|
||||
"* 最後，我們對所有字符進行獨熱編碼，並執行位移和 `<eos>` 附加操作。我們很快就會了解為什麼需要獨熱編碼的字符\n",
|
||||
"\n",
|
||||
"然而，這個函數是 **Pythonic** 的，也就是說它無法自動轉換為 Tensorflow 的計算圖。如果我們直接在 `Dataset.map` 函數中使用這個函數，會出現錯誤。我們需要使用 `py_function` 包裝器來封裝這個 Pythonic 調用：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 6,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"def title_batch_fn(x):\n",
|
||||
" x = x['title']\n",
|
||||
" a,b = tf.py_function(title_batch,inp=[x],Tout=(tf.float32,tf.float32))\n",
|
||||
" return a,b"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"> **注意**：區分 Pythonic 和 Tensorflow 的轉換函數可能看起來有些複雜，您可能會疑惑為什麼我們不在將數據集傳遞給 `fit` 之前使用標準的 Python 函數進行轉換。雖然這確實是可行的，但使用 `Dataset.map` 有一個巨大的優勢，因為數據轉換管道是通過 Tensorflow 的計算圖執行的，這可以利用 GPU 的計算能力，並且減少了在 CPU 和 GPU 之間傳遞數據的需求。\n",
|
||||
"\n",
|
||||
"現在我們可以構建生成器網絡並開始訓練。它可以基於我們在上一單元中討論的任何循環單元（簡單、LSTM 或 GRU）。在我們的示例中，我們將使用 LSTM。\n",
|
||||
"\n",
|
||||
"由於網絡以字符作為輸入，且詞彙表的大小相對較小，我們不需要嵌入層，經過一次熱編碼的輸入可以直接進入 LSTM 單元。輸出層將是一個 `Dense` 分類器，它會將 LSTM 的輸出轉換為一次熱編碼的標記編號。\n",
|
||||
"\n",
|
||||
"此外，因為我們處理的是可變長度的序列，我們可以使用 `Masking` 層來創建一個掩碼，忽略字符串中填充的部分。這並不是絕對必要的，因為我們對超過 `<eos>` 標記的部分並不太感興趣，但我們會使用它來獲得一些使用這類型層的經驗。`input_shape` 將是 `(None, vocab_size)`，其中 `None` 表示可變長度的序列，而輸出形狀也是 `(None, vocab_size)`，正如您可以從 `summary` 中看到的那樣：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Model: \"sequential\"\n",
|
||||
"_________________________________________________________________\n",
|
||||
"Layer (type) Output Shape Param # \n",
|
||||
"=================================================================\n",
|
||||
"masking (Masking) (None, None, 84) 0 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"lstm (LSTM) (None, None, 128) 109056 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"dense (Dense) (None, None, 84) 10836 \n",
|
||||
"=================================================================\n",
|
||||
"Total params: 119,892\n",
|
||||
"Trainable params: 119,892\n",
|
||||
"Non-trainable params: 0\n",
|
||||
"_________________________________________________________________\n",
|
||||
"15000/15000 [==============================] - 229s 15ms/step - loss: 1.5385\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<tensorflow.python.keras.callbacks.History at 0x7fa40c1245e0>"
|
||||
]
|
||||
},
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model = keras.models.Sequential([\n",
|
||||
" keras.layers.Masking(input_shape=(None,vocab_size)),\n",
|
||||
" keras.layers.LSTM(128,return_sequences=True),\n",
|
||||
" keras.layers.Dense(vocab_size,activation='softmax')\n",
|
||||
"])\n",
|
||||
"\n",
|
||||
"model.summary()\n",
|
||||
"model.compile(loss='categorical_crossentropy')\n",
|
||||
"\n",
|
||||
"model.fit(ds_train.batch(8).map(title_batch_fn))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 生成輸出\n",
|
||||
"\n",
|
||||
"現在我們已經訓練了模型，接下來我們希望使用它來生成一些輸出。首先，我們需要一種方法來解碼由一系列標記數字表示的文本。為此，我們可以使用 `tokenizer.sequences_to_texts` 函數；然而，該函數在字元級標記化時效果不佳。因此，我們將從標記器中提取標記字典（稱為 `word_index`），建立一個反向映射，並撰寫自己的解碼函數：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 10,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"reverse_map = {val:key for key, val in tokenizer.word_index.items()}\n",
|
||||
"\n",
|
||||
"def decode(x):\n",
|
||||
" return ''.join([reverse_map[t] for t in x])"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在，我們開始進行生成。我們將以某個字串 `start` 作為起點，將其編碼為序列 `inp`，接著在每一步中呼叫我們的網路來推斷下一個字元。\n",
|
||||
"\n",
|
||||
"網路的輸出 `out` 是一個包含 `vocab_size` 元素的向量，代表每個標記的概率。我們可以使用 `argmax` 找出最可能的標記編號。然後，我們將這個字元附加到已生成的標記列表中，並繼續進行生成。這個生成一個字元的過程會重複執行 `size` 次，以生成所需的字元數量。如果在過程中遇到 `eos_token`，我們會提前終止生成。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 12,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"'Today #39;s lead to strike for the strike for the strike for the strike (AFP)'"
|
||||
]
|
||||
},
|
||||
"execution_count": 12,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"def generate(model,size=100,start='Today '):\n",
|
||||
" inp = tokenizer.texts_to_sequences([start])[0]\n",
|
||||
" chars = inp\n",
|
||||
" for i in range(size):\n",
|
||||
" out = model(tf.expand_dims(tf.one_hot(inp,vocab_size),0))[0][-1]\n",
|
||||
" nc = tf.argmax(out)\n",
|
||||
" if nc==eos_token:\n",
|
||||
" break\n",
|
||||
" chars.append(nc.numpy())\n",
|
||||
" inp = inp+[nc]\n",
|
||||
" return decode(chars)\n",
|
||||
" \n",
|
||||
"generate(model)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 在訓練期間抽樣輸出\n",
|
||||
"\n",
|
||||
"由於我們沒有任何像 *準確率* 這樣的有用指標，我們唯一能夠判斷模型是否有所改善的方法就是在訓練期間通過**抽樣**生成的字串來觀察。為了實現這一點，我們將使用**回調函數**，即可以傳遞給 `fit` 函數的函數，並在訓練過程中定期被調用。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 13,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Epoch 1/3\n",
|
||||
"15000/15000 [==============================] - 226s 15ms/step - loss: 1.2703\n",
|
||||
"Today #39;s a lead in the company for the strike\n",
|
||||
"Epoch 2/3\n",
|
||||
"15000/15000 [==============================] - 227s 15ms/step - loss: 1.2057\n",
|
||||
"Today #39;s the Market Service on Security Start (AP)\n",
|
||||
"Epoch 3/3\n",
|
||||
"15000/15000 [==============================] - 226s 15ms/step - loss: 1.1752\n",
|
||||
"Today #39;s a line on the strike to start for the start\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<tensorflow.python.keras.callbacks.History at 0x7fa40c74e3d0>"
|
||||
]
|
||||
},
|
||||
"execution_count": 13,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"sampling_callback = keras.callbacks.LambdaCallback(\n",
|
||||
" on_epoch_end = lambda batch, logs: print(generate(model))\n",
|
||||
")\n",
|
||||
"\n",
|
||||
"model.fit(ds_train.batch(8).map(title_batch_fn),callbacks=[sampling_callback],epochs=3)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"這個範例已經生成了一些相當不錯的文本，但仍有多種方式可以進一步改進：\n",
|
||||
"\n",
|
||||
"* **更多文本**。我們僅使用了標題作為任務內容，但您可能希望嘗試使用完整文本。請記住，RNN在處理長序列方面表現不佳，因此可以將文本拆分成較短的句子，或者始終在固定的序列長度（例如，`num_chars`，假設為256）上進行訓練。您可以嘗試將上述範例改為這種架構，並參考[官方 Keras 教學](https://keras.io/examples/generative/lstm_character_level_text_generation/)作為靈感。\n",
|
||||
"\n",
|
||||
"* **多層 LSTM**。嘗試使用2或3層的LSTM單元是有意義的。如我們在前一單元提到的，每層LSTM會從文本中提取特定的模式，而在字元級生成器的情況下，我們可以預期較低層的LSTM負責提取音節，而較高層則負責提取單詞及單詞組合。這可以通過向LSTM構造函數傳遞層數參數來簡單實現。\n",
|
||||
"\n",
|
||||
"* 您可能還希望嘗試使用**GRU單元**，看看哪種表現更好，以及嘗試**不同的隱藏層大小**。隱藏層過大可能導致過擬合（例如，網絡會學習精確的文本），而過小的隱藏層可能無法生成良好的結果。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 軟性文本生成與溫度\n",
|
||||
"\n",
|
||||
"在之前 `generate` 的定義中，我們總是選擇機率最高的字元作為生成文本的下一個字元。這導致文本經常在相同的字元序列之間不斷「循環」，如下例所示：\n",
|
||||
"```\n",
|
||||
"today of the second the company and a second the company ...\n",
|
||||
"```\n",
|
||||
"\n",
|
||||
"然而，如果我們觀察下一個字元的機率分佈，可能會發現幾個最高機率之間的差距並不大，例如一個字元的機率是 0.2，另一個是 0.19，等等。例如，在尋找序列 '*play*' 的下一個字元時，下一個字元可能同樣有可能是空格，或者是 **e**（如單詞 *player* 中的情況）。\n",
|
||||
"\n",
|
||||
"這讓我們得出一個結論：選擇機率最高的字元並不總是「公平」的，因為選擇第二高的字元仍然可能生成有意義的文本。更明智的做法是從網路輸出的機率分佈中**抽樣**字元。\n",
|
||||
"\n",
|
||||
"這種抽樣可以使用 `np.multinomial` 函數來完成，該函數實現了所謂的**多項分佈**。下面定義了一個實現這種**軟性**文本生成的函數：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 33,
|
||||
"metadata": {
|
||||
"scrolled": true
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"\n",
|
||||
"--- Temperature = 0.3\n",
|
||||
"Today #39;s strike #39; to start at the store return\n",
|
||||
"On Sunday PO to Be Data Profit Up (Reuters)\n",
|
||||
"Moscow, SP wins straight to the Microsoft #39;s control of the space start\n",
|
||||
"President olding of the blast start for the strike to pay <b>...</b>\n",
|
||||
"Little red riding hood ficed to the spam countered in European <b>...</b>\n",
|
||||
"\n",
|
||||
"--- Temperature = 0.8\n",
|
||||
"Today countie strikes ryder missile faces food market blut\n",
|
||||
"On Sunday collores lose-toppy of sale of Bullment in <b>...</b>\n",
|
||||
"Moscow, IBM Diffeiting in Afghan Software Hotels (Reuters)\n",
|
||||
"President Ol Luster for Profit Peaced Raised (AP)\n",
|
||||
"Little red riding hood dace on depart talks #39; bank up\n",
|
||||
"\n",
|
||||
"--- Temperature = 1.0\n",
|
||||
"Today wits House buiting debate fixes #39; supervice stake again\n",
|
||||
"On Sunday arling digital poaching In for level\n",
|
||||
"Moscow, DS Up 7, Top Proble Protest Caprey Mamarian Strike\n",
|
||||
"President teps help of roubler stepted lessabul-Dhalitics (AFP)\n",
|
||||
"Little red riding hood signs on cash in Carter-youb\n",
|
||||
"\n",
|
||||
"--- Temperature = 1.3\n",
|
||||
"Today wits flawer ro, pSIA figat's co DroftwavesIs Talo up\n",
|
||||
"On Sunday hround elitwing wint EU Powerburlinetien\n",
|
||||
"Moscow, Bazz #39;s sentries olymen winnelds' next for Olympite Huc?\n",
|
||||
"President lost securitys from power Elections in Smiltrials\n",
|
||||
"Little red riding hood vides profit, exponituity, profitmainalist-at said listers\n",
|
||||
"\n",
|
||||
"--- Temperature = 1.8\n",
|
||||
"Today #39;It: He deat: N.KA Asside\n",
|
||||
"On Sunday i arry Par aldeup patient Wo stele1\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"ename": "KeyError",
|
||||
"evalue": "0",
|
||||
"output_type": "error",
|
||||
"traceback": [
|
||||
"\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
|
||||
"\u001b[0;31mKeyError\u001b[0m Traceback (most recent call last)",
|
||||
"\u001b[0;32m<ipython-input-33-db32367a0feb>\u001b[0m in \u001b[0;36m<module>\u001b[0;34m\u001b[0m\n\u001b[1;32m 18\u001b[0m \u001b[0mprint\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34mf\"\\n--- Temperature = {i}\"\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 19\u001b[0m \u001b[0;32mfor\u001b[0m \u001b[0mj\u001b[0m \u001b[0;32min\u001b[0m \u001b[0mrange\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;36m5\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 20\u001b[0;31m \u001b[0mprint\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mgenerate_soft\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mmodel\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0msize\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0;36m300\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0mstart\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0mwords\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0mj\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0mtemperature\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0mi\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m",
|
||||
"\u001b[0;32m<ipython-input-33-db32367a0feb>\u001b[0m in \u001b[0;36mgenerate_soft\u001b[0;34m(model, size, start, temperature)\u001b[0m\n\u001b[1;32m 11\u001b[0m \u001b[0mchars\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mappend\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mnc\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 12\u001b[0m \u001b[0minp\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0minp\u001b[0m\u001b[0;34m+\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0mnc\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 13\u001b[0;31m \u001b[0;32mreturn\u001b[0m \u001b[0mdecode\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mchars\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m 14\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 15\u001b[0m \u001b[0mwords\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0;34m[\u001b[0m\u001b[0;34m'Today '\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m'On Sunday '\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m'Moscow, '\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m'President '\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m'Little red riding hood '\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
|
||||
"\u001b[0;32m<ipython-input-10-3f5fa6130b1d>\u001b[0m in \u001b[0;36mdecode\u001b[0;34m(x)\u001b[0m\n\u001b[1;32m 2\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 3\u001b[0m \u001b[0;32mdef\u001b[0m \u001b[0mdecode\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mx\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m----> 4\u001b[0;31m \u001b[0;32mreturn\u001b[0m \u001b[0;34m''\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mjoin\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0mreverse_map\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0mt\u001b[0m\u001b[0;34m]\u001b[0m \u001b[0;32mfor\u001b[0m \u001b[0mt\u001b[0m \u001b[0;32min\u001b[0m \u001b[0mx\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m",
|
||||
"\u001b[0;32m<ipython-input-10-3f5fa6130b1d>\u001b[0m in \u001b[0;36m<listcomp>\u001b[0;34m(.0)\u001b[0m\n\u001b[1;32m 2\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 3\u001b[0m \u001b[0;32mdef\u001b[0m \u001b[0mdecode\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mx\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m----> 4\u001b[0;31m \u001b[0;32mreturn\u001b[0m \u001b[0;34m''\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mjoin\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0mreverse_map\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0mt\u001b[0m\u001b[0;34m]\u001b[0m \u001b[0;32mfor\u001b[0m \u001b[0mt\u001b[0m \u001b[0;32min\u001b[0m \u001b[0mx\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m",
|
||||
"\u001b[0;31mKeyError\u001b[0m: 0"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"def generate_soft(model,size=100,start='Today ',temperature=1.0):\n",
|
||||
" inp = tokenizer.texts_to_sequences([start])[0]\n",
|
||||
" chars = inp\n",
|
||||
" for i in range(size):\n",
|
||||
" out = model(tf.expand_dims(tf.one_hot(inp,vocab_size),0))[0][-1]\n",
|
||||
" probs = tf.exp(tf.math.log(out)/temperature).numpy().astype(np.float64)\n",
|
||||
" probs = probs/np.sum(probs)\n",
|
||||
" nc = np.argmax(np.random.multinomial(1,probs,1))\n",
|
||||
" if nc==eos_token:\n",
|
||||
" break\n",
|
||||
" chars.append(nc)\n",
|
||||
" inp = inp+[nc]\n",
|
||||
" return decode(chars)\n",
|
||||
"\n",
|
||||
"words = ['Today ','On Sunday ','Moscow, ','President ','Little red riding hood ']\n",
|
||||
" \n",
|
||||
"for i in [0.3,0.8,1.0,1.3,1.8]:\n",
|
||||
" print(f\"\\n--- Temperature = {i}\")\n",
|
||||
" for j in range(5):\n",
|
||||
" print(generate_soft(model,size=300,start=words[j],temperature=i))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們引入了一個名為 **溫度** 的參數，用於指示我們應該多大程度地堅持最高概率。如果溫度為 1.0，我們進行公平的多項式抽樣，而當溫度趨於無窮大時，所有概率變得相等，我們隨機選擇下一個字符。在下面的例子中，我們可以觀察到當溫度增加過多時，文本變得毫無意義，而當溫度接近 0 時，則類似於「循環」的硬生成文本。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件使用 AI 翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。應以原文文件作為權威來源。對於關鍵資訊，建議尋求專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或誤讀概不負責。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"interpreter": {
|
||||
"hash": "16af2a8bbb083ea23e5e41c7f5787656b2ce26968575d8763f2c4b17f9cd711f"
|
||||
},
|
||||
"kernelspec": {
|
||||
"display_name": "Python 3.8.12 ('py38')",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.8.12"
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "9fbb7d5fda708537649f71f5f646fcde",
|
||||
"translation_date": "2025-08-31T10:31:48+00:00",
|
||||
"source_file": "lessons/5-NLP/17-GenerativeNetworks/GenerativeTF.ipynb",
|
||||
"language_code": "tw"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 4
|
||||
}
|
||||
|
|
@ -0,0 +1,353 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 注意力機制與Transformer\n",
|
||||
"\n",
|
||||
"循環神經網路（RNN）的一個主要缺點是序列中的所有詞對結果的影響相同。這導致標準的LSTM編碼器-解碼器模型在處理序列到序列任務（如命名實體識別和機器翻譯）時表現不佳。實際上，輸入序列中的某些特定詞往往對輸出序列的影響更大。\n",
|
||||
"\n",
|
||||
"考慮一個序列到序列的模型，例如機器翻譯。這種模型由兩個循環神經網路實現，其中一個網路（**編碼器**）將輸入序列壓縮成隱藏狀態，另一個網路（**解碼器**）將該隱藏狀態展開為翻譯結果。這種方法的問題在於，網路的最終狀態很難記住句子的開頭部分，從而導致模型在處理長句子時質量較差。\n",
|
||||
"\n",
|
||||
"**注意力機制**提供了一種方法，能夠對每個輸入向量對RNN每個輸出預測的上下文影響進行加權。其實現方式是通過在輸入RNN的中間狀態和輸出RNN之間創建捷徑。這樣，在生成輸出符號$y_t$時，我們會考慮所有輸入的隱藏狀態$h_i$，並賦予不同的權重係數$\\alpha_{t,i}$。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"*帶有加性注意力機制的編碼器-解碼器模型，來自 [Bahdanau et al., 2015](https://arxiv.org/pdf/1409.0473.pdf)，引用自[這篇博客文章](https://lilianweng.github.io/lil-log/2018/06/24/attention-attention.html)*\n",
|
||||
"\n",
|
||||
"注意力矩陣$\\{\\alpha_{i,j}\\}$表示某些輸入詞在生成輸出序列中特定詞時的影響程度。以下是這樣一個矩陣的示例：\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"*圖片來自 [Bahdanau et al., 2015](https://arxiv.org/pdf/1409.0473.pdf)（圖3）*\n",
|
||||
"\n",
|
||||
"注意力機制是當前或接近當前自然語言處理技術水平的關鍵。儘管如此，添加注意力機制會大幅增加模型參數的數量，這導致了RNN的擴展問題。RNN的一個關鍵限制是其循環特性使得訓練過程難以批量化和並行化。在RNN中，序列的每個元素都需要按順序處理，這意味著它無法輕易並行化。\n",
|
||||
"\n",
|
||||
"注意力機制的採用結合這一限制，促使了如今的Transformer模型的誕生，這些模型（如BERT和OpenGPT3）已成為當前的技術標杆。\n",
|
||||
"\n",
|
||||
"## Transformer模型\n",
|
||||
"\n",
|
||||
"與將每個先前預測的上下文傳遞到下一個評估步驟不同，**Transformer模型**使用**位置編碼**和注意力機制來捕捉給定輸入在提供的文本窗口內的上下文。下圖展示了如何通過位置編碼和注意力機制在給定窗口內捕捉上下文。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"由於每個輸入位置可以獨立映射到每個輸出位置，Transformer比RNN更容易並行化，這使得構建更大、更具表達力的語言模型成為可能。每個注意力頭可以用來學習詞與詞之間的不同關係，從而改進下游的自然語言處理任務。\n",
|
||||
"\n",
|
||||
"**BERT**（Bidirectional Encoder Representations from Transformers）是一個非常大的多層Transformer網路，*BERT-base*有12層，*BERT-large*有24層。該模型首先在大規模文本數據（維基百科+書籍）上進行無監督預訓練（預測句子中的被遮蔽詞）。在預訓練過程中，模型吸收了大量的語言理解能力，這些能力可以通過微調其他數據集來利用。這個過程稱為**遷移學習**。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"Transformer架構有許多變體，包括BERT、DistilBERT、BigBird、OpenGPT3等，這些模型都可以進行微調。[HuggingFace套件](https://github.com/huggingface/)提供了使用PyTorch訓練這些架構的資源庫。\n",
|
||||
"\n",
|
||||
"## 使用BERT進行文本分類\n",
|
||||
"\n",
|
||||
"讓我們看看如何使用預訓練的BERT模型來解決我們的傳統任務：序列分類。我們將對原始的AG News數據集進行分類。\n",
|
||||
"\n",
|
||||
"首先，讓我們加載HuggingFace庫和數據集：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 10,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Loading dataset...\n",
|
||||
"Building vocab...\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"import torch\n",
|
||||
"import torchtext\n",
|
||||
"from torchnlp import *\n",
|
||||
"import transformers\n",
|
||||
"train_dataset, test_dataset, classes, vocab = load_dataset()\n",
|
||||
"vocab_len = len(vocab)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"由於我們將使用預訓練的 BERT 模型，因此需要使用特定的分詞器。首先，我們將載入與預訓練 BERT 模型相關聯的分詞器。\n",
|
||||
"\n",
|
||||
"HuggingFace 庫包含一個預訓練模型的存儲庫，您只需在 `from_pretrained` 函數中指定模型名稱作為參數即可使用。所有模型所需的二進制文件將自動下載。\n",
|
||||
"\n",
|
||||
"然而，有時您可能需要載入自己的模型，在這種情況下，您可以指定包含所有相關文件的目錄，包括分詞器的參數、模型參數的 `config.json` 文件、二進制權重等。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 11,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"# To load the model from Internet repository using model name. \n",
|
||||
"# Use this if you are running from your own copy of the notebooks\n",
|
||||
"bert_model = 'bert-base-uncased' \n",
|
||||
"\n",
|
||||
"# To load the model from the directory on disk. Use this for Microsoft Learn module, because we have\n",
|
||||
"# prepared all required files for you.\n",
|
||||
"bert_model = './bert'\n",
|
||||
"\n",
|
||||
"tokenizer = transformers.BertTokenizer.from_pretrained(bert_model)\n",
|
||||
"\n",
|
||||
"MAX_SEQ_LEN = 128\n",
|
||||
"PAD_INDEX = tokenizer.convert_tokens_to_ids(tokenizer.pad_token)\n",
|
||||
"UNK_INDEX = tokenizer.convert_tokens_to_ids(tokenizer.unk_token)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"`tokenizer` 物件包含可直接用於編碼文本的 `encode` 函數：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 15,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"[101, 1052, 22123, 2953, 2818, 2003, 1037, 2307, 7705, 2005, 17953, 2361, 102]"
|
||||
]
|
||||
},
|
||||
"execution_count": 15,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"tokenizer.encode('PyTorch is a great framework for NLP')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"然後，讓我們建立在訓練期間用於訪問數據的迭代器。由於 BERT 使用其自己的編碼函數，我們需要定義一個類似於之前定義的 `padify` 的填充函數：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"def pad_bert(b):\n",
|
||||
" # b is the list of tuples of length batch_size\n",
|
||||
" # - first element of a tuple = label, \n",
|
||||
" # - second = feature (text sequence)\n",
|
||||
" # build vectorized sequence\n",
|
||||
" v = [tokenizer.encode(x[1]) for x in b]\n",
|
||||
" # compute max length of a sequence in this minibatch\n",
|
||||
" l = max(map(len,v))\n",
|
||||
" return ( # tuple of two tensors - labels and features\n",
|
||||
" torch.LongTensor([t[0] for t in b]),\n",
|
||||
" torch.stack([torch.nn.functional.pad(torch.tensor(t),(0,l-len(t)),mode='constant',value=0) for t in v])\n",
|
||||
" )\n",
|
||||
"\n",
|
||||
"train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=8, collate_fn=pad_bert, shuffle=True)\n",
|
||||
"test_loader = torch.utils.data.DataLoader(test_dataset, batch_size=8, collate_fn=pad_bert)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"在我們的案例中，我們將使用名為 `bert-base-uncased` 的預訓練 BERT 模型。讓我們使用 `BertForSequenceClassfication` 套件來加載模型。這確保了我們的模型已經具備分類所需的架構，包括最終的分類器。您會看到一條警告訊息，指出最終分類器的權重尚未初始化，並且模型需要進行預訓練——這完全沒問題，因為這正是我們即將進行的操作！\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 9,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stderr",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Some weights of the model checkpoint at ./bert were not used when initializing BertForSequenceClassification: ['cls.predictions.bias', 'cls.predictions.transform.dense.weight', 'cls.predictions.transform.dense.bias', 'cls.predictions.decoder.weight', 'cls.seq_relationship.weight', 'cls.seq_relationship.bias', 'cls.predictions.transform.LayerNorm.weight', 'cls.predictions.transform.LayerNorm.bias']\n",
|
||||
"- This IS expected if you are initializing BertForSequenceClassification from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).\n",
|
||||
"- This IS NOT expected if you are initializing BertForSequenceClassification from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).\n",
|
||||
"Some weights of BertForSequenceClassification were not initialized from the model checkpoint at ./bert and are newly initialized: ['classifier.weight', 'classifier.bias']\n",
|
||||
"You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model = transformers.BertForSequenceClassification.from_pretrained(bert_model,num_labels=4).to(device)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在我們準備開始訓練了！由於 BERT 已經是預訓練模型，我們希望從較小的學習率開始，以免破壞初始權重。\n",
|
||||
"\n",
|
||||
"所有的繁重工作都由 `BertForSequenceClassification` 模型完成。當我們將訓練數據傳入模型時，它會返回損失值和輸入小批量的網絡輸出。我們使用損失值進行參數優化（`loss.backward()` 執行反向傳播），並使用 `out` 計算訓練準確率，方法是將獲得的標籤 `labs`（使用 `argmax` 計算）與預期的 `labels` 進行比較。\n",
|
||||
"\n",
|
||||
"為了控制訓練過程，我們會在多次迭代中累積損失值和準確率，並在每 `report_freq` 個訓練週期後打印它們。\n",
|
||||
"\n",
|
||||
"這次訓練可能會花費相當長的時間，因此我們限制了迭代次數。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 6,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Loss = 1.1254194641113282, Accuracy = 0.585\n",
|
||||
"Loss = 0.6194715118408203, Accuracy = 0.83\n",
|
||||
"Loss = 0.46665248870849607, Accuracy = 0.8475\n",
|
||||
"Loss = 0.4309701919555664, Accuracy = 0.8575\n",
|
||||
"Loss = 0.35427074432373046, Accuracy = 0.8825\n",
|
||||
"Loss = 0.3306886291503906, Accuracy = 0.8975\n",
|
||||
"Loss = 0.30340143203735354, Accuracy = 0.8975\n",
|
||||
"Loss = 0.26139299392700194, Accuracy = 0.915\n",
|
||||
"Loss = 0.26708646774291994, Accuracy = 0.9225\n",
|
||||
"Loss = 0.3667240524291992, Accuracy = 0.8675\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"optimizer = torch.optim.Adam(model.parameters(), lr=2e-5)\n",
|
||||
"\n",
|
||||
"report_freq = 50\n",
|
||||
"iterations = 500 # make this larger to train for longer time!\n",
|
||||
"\n",
|
||||
"model.train()\n",
|
||||
"\n",
|
||||
"i,c = 0,0\n",
|
||||
"acc_loss = 0\n",
|
||||
"acc_acc = 0\n",
|
||||
"\n",
|
||||
"for labels,texts in train_loader:\n",
|
||||
" labels = labels.to(device)-1 # get labels in the range 0-3 \n",
|
||||
" texts = texts.to(device)\n",
|
||||
" loss, out = model(texts, labels=labels)[:2]\n",
|
||||
" labs = out.argmax(dim=1)\n",
|
||||
" acc = torch.mean((labs==labels).type(torch.float32))\n",
|
||||
" optimizer.zero_grad()\n",
|
||||
" loss.backward()\n",
|
||||
" optimizer.step()\n",
|
||||
" acc_loss += loss\n",
|
||||
" acc_acc += acc\n",
|
||||
" i+=1\n",
|
||||
" c+=1\n",
|
||||
" if i%report_freq==0:\n",
|
||||
" print(f\"Loss = {acc_loss.item()/c}, Accuracy = {acc_acc.item()/c}\")\n",
|
||||
" c = 0\n",
|
||||
" acc_loss = 0\n",
|
||||
" acc_acc = 0\n",
|
||||
" iterations-=1\n",
|
||||
" if not iterations:\n",
|
||||
" break"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"你可以看到（尤其是當你增加迭代次數並等待足夠長的時間），BERT 的分類能夠提供相當不錯的準確率！這是因為 BERT 已經對語言的結構有相當好的理解，我們只需要微調最終的分類器即可。然而，由於 BERT 是一個大型模型，整個訓練過程需要很長的時間，並且需要強大的計算能力！（GPU，最好是多個 GPU）。\n",
|
||||
"\n",
|
||||
"> **Note:** 在我們的範例中，我們使用的是最小的預訓練 BERT 模型之一。還有更大的模型可能會產生更好的結果。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 評估模型表現\n",
|
||||
"\n",
|
||||
"現在我們可以在測試數據集上評估模型的表現。評估循環與訓練循環非常相似，但我們不能忘記通過調用 `model.eval()` 將模型切換到評估模式。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 10,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Final accuracy: 0.9047029702970297\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model.eval()\n",
|
||||
"iterations = 100\n",
|
||||
"acc = 0\n",
|
||||
"i = 0\n",
|
||||
"for labels,texts in test_loader:\n",
|
||||
" labels = labels.to(device)-1 \n",
|
||||
" texts = texts.to(device)\n",
|
||||
" _, out = model(texts, labels=labels)[:2]\n",
|
||||
" labs = out.argmax(dim=1)\n",
|
||||
" acc += torch.mean((labs==labels).type(torch.float32))\n",
|
||||
" i+=1\n",
|
||||
" if i>iterations: break\n",
|
||||
" \n",
|
||||
"print(f\"Final accuracy: {acc.item()/i}\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 重點\n",
|
||||
"\n",
|
||||
"在本單元中，我們已經看到從 **transformers** 庫中取用預訓練語言模型並將其適配於我們的文本分類任務是多麼簡單。同樣地，BERT 模型也可以用於實體抽取、問答系統以及其他 NLP 任務。\n",
|
||||
"\n",
|
||||
"Transformer 模型代表了 NLP 領域的最新技術，在大多數情況下，當實現自定義 NLP 解決方案時，它應該是您開始實驗的首選解決方案。然而，如果您希望構建更高級的神經網絡模型，理解本模組中討論的循環神經網絡的基本原理是非常重要的。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件已使用 AI 翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。儘管我們致力於提供準確的翻譯，請注意自動翻譯可能包含錯誤或不準確之處。原始文件的母語版本應被視為權威來源。對於關鍵資訊，建議使用專業人工翻譯。我們對因使用此翻譯而引起的任何誤解或錯誤解釋不承擔責任。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"kernelspec": {
|
||||
"display_name": "py37_pytorch",
|
||||
"language": "python",
|
||||
"name": "conda-env-py37_pytorch-py"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.7.7"
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "753865967678a92dbce7d7efbd36d980",
|
||||
"translation_date": "2025-08-31T10:39:09+00:00",
|
||||
"source_file": "lessons/5-NLP/18-Transformers/TransformersPyTorch.ipynb",
|
||||
"language_code": "tw"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 4
|
||||
}
|
||||
|
|
@ -0,0 +1,825 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"# 注意力機制與Transformer模型\n",
|
||||
"\n",
|
||||
"循環神經網路（RNN）的一個主要缺點是序列中的所有詞對結果的影響相同。這導致標準的LSTM編碼器-解碼器模型在處理序列到序列任務（如命名實體識別和機器翻譯）時表現不佳。實際上，輸入序列中的某些特定詞往往對輸出序列的影響更大。\n",
|
||||
"\n",
|
||||
"考慮一個序列到序列的模型，例如機器翻譯。這種模型由兩個循環神經網路實現，其中一個網路（**編碼器**）將輸入序列壓縮成隱藏狀態，另一個網路（**解碼器**）將該隱藏狀態展開為翻譯結果。這種方法的問題在於，網路的最終狀態很難記住句子的開頭部分，從而導致模型在處理長句子時質量較差。\n",
|
||||
"\n",
|
||||
"**注意力機制**提供了一種方法，能夠對每個輸入向量對RNN每個輸出預測的上下文影響進行加權。其實現方式是通過在輸入RNN的中間狀態和輸出RNN之間創建捷徑。這樣，在生成輸出符號$y_t$時，我們會考慮所有輸入的隱藏狀態$h_i$，並賦予不同的權重係數$\\alpha_{t,i}$。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"*來自 [Bahdanau et al., 2015](https://arxiv.org/pdf/1409.0473.pdf) 的加性注意力機制編碼器-解碼器模型，圖片引用自[這篇博客文章](https://lilianweng.github.io/lil-log/2018/06/24/attention-attention.html)*\n",
|
||||
"\n",
|
||||
"注意力矩陣$\\{\\alpha_{i,j}\\}$表示某些輸入詞在生成輸出序列中特定詞時所起的作用程度。以下是這樣一個矩陣的示例：\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"*圖片取自 [Bahdanau et al., 2015](https://arxiv.org/pdf/1409.0473.pdf)（圖3）*\n",
|
||||
"\n",
|
||||
"注意力機制是當前或接近當前自然語言處理（NLP）技術水平的核心。儘管如此，添加注意力機制會大幅增加模型參數的數量，這導致了RNN的擴展問題。RNN的一個關鍵限制是其循環特性使得訓練過程難以批量化和並行化。在RNN中，序列的每個元素都需要按順序處理，這意味著它無法輕易並行化。\n",
|
||||
"\n",
|
||||
"注意力機制的採用結合了這一限制，促使了如今我們熟知並使用的Transformer模型的誕生，從BERT到OpenGPT3，這些模型代表了當前的技術水平。\n",
|
||||
"\n",
|
||||
"## Transformer模型\n",
|
||||
"\n",
|
||||
"與將每次預測的上下文傳遞到下一個評估步驟不同，**Transformer模型**使用**位置編碼**和**注意力機制**來捕捉給定輸入在提供的文本窗口內的上下文。下圖展示了如何通過位置編碼和注意力機制在給定窗口內捕捉上下文。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"由於每個輸入位置可以獨立映射到每個輸出位置，Transformer模型比RNN更容易並行化，這使得構建更大、更具表達力的語言模型成為可能。每個注意力頭可以用來學習詞與詞之間的不同關係，從而改進下游的自然語言處理任務。\n",
|
||||
"\n",
|
||||
"## 構建簡單的Transformer模型\n",
|
||||
"\n",
|
||||
"Keras中並未內建Transformer層，但我們可以自行構建。與之前一樣，我們將專注於AG News數據集的文本分類，但值得一提的是，Transformer模型在更困難的NLP任務中表現最佳。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 1,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import tensorflow as tf\n",
|
||||
"from tensorflow import keras\n",
|
||||
"import tensorflow_datasets as tfds\n",
|
||||
"import numpy as np\n",
|
||||
"\n",
|
||||
"ds_train, ds_test = tfds.load('ag_news_subset').values()\n",
|
||||
"\n",
|
||||
"def extract_text(x):\n",
|
||||
" return x['title']+' '+x['description']\n",
|
||||
"\n",
|
||||
"def tupelize(x):\n",
|
||||
" return (extract_text(x),x['label'])"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"新的 Keras 層應該繼承 `Layer` 類別，並實現 `call` 方法。我們先從 **位置嵌入** 層開始。我們將使用[官方 Keras 文件中的一些代碼](https://keras.io/examples/nlp/text_classification_with_transformer/)。我們假設我們將所有輸入序列填充到長度 `maxlen`。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 2,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"class TokenAndPositionEmbedding(keras.layers.Layer):\n",
|
||||
" def __init__(self, maxlen, vocab_size, embed_dim):\n",
|
||||
" super(TokenAndPositionEmbedding, self).__init__()\n",
|
||||
" self.token_emb = keras.layers.Embedding(input_dim=vocab_size, output_dim=embed_dim)\n",
|
||||
" self.pos_emb = keras.layers.Embedding(input_dim=maxlen, output_dim=embed_dim)\n",
|
||||
" self.maxlen = maxlen\n",
|
||||
"\n",
|
||||
" def call(self, x):\n",
|
||||
" maxlen = self.maxlen\n",
|
||||
" positions = tf.range(start=0, limit=maxlen, delta=1)\n",
|
||||
" positions = self.pos_emb(positions)\n",
|
||||
" x = self.token_emb(x)\n",
|
||||
" return x+positions"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"這一層包含兩個 `Embedding` 層：一個用於嵌入標記（以我們之前討論過的方式），另一個用於嵌入標記位置。標記位置是通過使用 `tf.range` 從 0 到 `maxlen` 創建的一系列自然數，然後通過嵌入層處理。兩個生成的嵌入向量隨後相加，產生輸入的基於位置嵌入的表示，其形狀為 `maxlen`$\\times$`embed_dim`。\n",
|
||||
"\n",
|
||||
"現在，我們來實現 transformer block。它將接收之前定義的嵌入層的輸出：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"class TransformerBlock(keras.layers.Layer):\n",
|
||||
" def __init__(self, embed_dim, num_heads, ff_dim, rate=0.1):\n",
|
||||
" super(TransformerBlock, self).__init__()\n",
|
||||
" self.att = keras.layers.MultiHeadAttention(num_heads=num_heads, key_dim=embed_dim, name='attn')\n",
|
||||
" self.ffn = keras.Sequential(\n",
|
||||
" [keras.layers.Dense(ff_dim, activation=\"relu\"), keras.layers.Dense(embed_dim),]\n",
|
||||
" )\n",
|
||||
" self.layernorm1 = keras.layers.LayerNormalization(epsilon=1e-6)\n",
|
||||
" self.layernorm2 = keras.layers.LayerNormalization(epsilon=1e-6)\n",
|
||||
" self.dropout1 = keras.layers.Dropout(rate)\n",
|
||||
" self.dropout2 = keras.layers.Dropout(rate)\n",
|
||||
"\n",
|
||||
" def call(self, inputs, training):\n",
|
||||
" attn_output = self.att(inputs, inputs)\n",
|
||||
" attn_output = self.dropout1(attn_output, training=training)\n",
|
||||
" out1 = self.layernorm1(inputs + attn_output)\n",
|
||||
" ffn_output = self.ffn(out1)\n",
|
||||
" ffn_output = self.dropout2(ffn_output, training=training)\n",
|
||||
" return self.layernorm2(out1 + ffn_output)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"Transformer 將 `MultiHeadAttention` 應用於帶有位置編碼的輸入，生成維度為 `maxlen`$\\times$`embed_dim` 的注意力向量，然後與輸入混合，並使用 `LayerNormalization` 進行正規化。\n",
|
||||
"\n",
|
||||
"> **注意**: `LayerNormalization` 與在本學習路徑 *計算機視覺* 部分中討論的 `BatchNormalization` 類似，但它針對每個訓練樣本獨立地正規化前一層的輸出，將其範圍調整到 [-1..1]。\n",
|
||||
"\n",
|
||||
"該層的輸出隨後通過 `Dense` 網絡（在我們的例子中是兩層感知機），結果再加到最終輸出中（該輸出再次經過正規化處理）。 \n",
|
||||
"\n",
|
||||
"現在，我們準備定義完整的 Transformer 模型：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Model: \"sequential_1\"\n",
|
||||
"_________________________________________________________________\n",
|
||||
"Layer (type) Output Shape Param # \n",
|
||||
"=================================================================\n",
|
||||
"text_vectorization (TextVect (None, 256) 0 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"token_and_position_embedding (None, 256, 32) 648192 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"transformer_block (Transform (None, 256, 32) 10656 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"global_average_pooling1d (Gl (None, 32) 0 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"dropout_2 (Dropout) (None, 32) 0 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"dense_2 (Dense) (None, 20) 660 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"dropout_3 (Dropout) (None, 20) 0 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"dense_3 (Dense) (None, 4) 84 \n",
|
||||
"=================================================================\n",
|
||||
"Total params: 659,592\n",
|
||||
"Trainable params: 659,592\n",
|
||||
"Non-trainable params: 0\n",
|
||||
"_________________________________________________________________\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"embed_dim = 32 # Embedding size for each token\n",
|
||||
"num_heads = 2 # Number of attention heads\n",
|
||||
"ff_dim = 32 # Hidden layer size in feed forward network inside transformer\n",
|
||||
"maxlen = 256\n",
|
||||
"vocab_size = 20000\n",
|
||||
"\n",
|
||||
"model = keras.models.Sequential([\n",
|
||||
" keras.layers.experimental.preprocessing.TextVectorization(max_tokens=vocab_size,output_sequence_length=maxlen, input_shape=(1,)),\n",
|
||||
" TokenAndPositionEmbedding(maxlen, vocab_size, embed_dim),\n",
|
||||
" TransformerBlock(embed_dim, num_heads, ff_dim),\n",
|
||||
" keras.layers.GlobalAveragePooling1D(),\n",
|
||||
" keras.layers.Dropout(0.1),\n",
|
||||
" keras.layers.Dense(20, activation=\"relu\"),\n",
|
||||
" keras.layers.Dropout(0.1),\n",
|
||||
" keras.layers.Dense(4, activation=\"softmax\")\n",
|
||||
"])\n",
|
||||
"\n",
|
||||
"model.summary()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Training tokenizer\n",
|
||||
"938/938 [==============================] - 45s 39ms/step - loss: 0.4978 - acc: 0.8068 - val_loss: 0.2808 - val_acc: 0.9124\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<tensorflow.python.keras.callbacks.History at 0x7f9c2427a0d0>"
|
||||
]
|
||||
},
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"print('Training tokenizer')\n",
|
||||
"model.layers[0].adapt(ds_train.map(extract_text))\n",
|
||||
"model.compile(loss='sparse_categorical_crossentropy',metrics=['acc'], optimizer='adam')\n",
|
||||
"model.fit(ds_train.map(tupelize).batch(128),validation_data=ds_test.map(tupelize).batch(128))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## BERT Transformer 模型\n",
|
||||
"\n",
|
||||
"**BERT**（Bidirectional Encoder Representations from Transformers，雙向編碼器表示）是一個非常大型的多層 Transformer 網絡，*BERT-base* 有 12 層，*BERT-large* 則有 24 層。該模型首先在大規模文本數據（維基百科 + 書籍）上進行無監督訓練（預測句子中被遮蔽的詞語）。在預訓練過程中，模型吸收了大量的語言理解能力，隨後可以通過微調其他數據集來加以利用。這個過程被稱為 **遷移學習**。\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"Transformer 架構有許多變體，包括 BERT、DistilBERT、BigBird、OpenGPT3 等，這些模型都可以進行微調。\n",
|
||||
"\n",
|
||||
"接下來，我們來看看如何使用預訓練的 BERT 模型來解決我們傳統的序列分類問題。我們將借用[官方文檔](https://www.tensorflow.org/text/tutorials/classify_text_with_bert)中的一些想法和代碼。\n",
|
||||
"\n",
|
||||
"為了加載預訓練模型，我們將使用 **Tensorflow hub**。首先，讓我們加載 BERT 專用的向量化工具：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 1,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"ename": "ModuleNotFoundError",
|
||||
"evalue": "No module named 'tensorflow_text'",
|
||||
"output_type": "error",
|
||||
"traceback": [
|
||||
"\u001b[1;31m---------------------------------------------------------------------------\u001b[0m",
|
||||
"\u001b[1;31mModuleNotFoundError\u001b[0m Traceback (most recent call last)",
|
||||
"\u001b[1;32m~\\AppData\\Local\\Temp/ipykernel_41180/4216669875.py\u001b[0m in \u001b[0;36m<module>\u001b[1;34m\u001b[0m\n\u001b[1;32m----> 1\u001b[1;33m \u001b[1;32mimport\u001b[0m \u001b[0mtensorflow_text\u001b[0m\u001b[1;33m\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0m\u001b[0;32m 2\u001b[0m \u001b[1;32mimport\u001b[0m \u001b[0mtensorflow_hub\u001b[0m \u001b[1;32mas\u001b[0m \u001b[0mhub\u001b[0m\u001b[1;33m\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0;32m 3\u001b[0m \u001b[0mvectorizer\u001b[0m \u001b[1;33m=\u001b[0m \u001b[0mhub\u001b[0m\u001b[1;33m.\u001b[0m\u001b[0mKerasLayer\u001b[0m\u001b[1;33m(\u001b[0m\u001b[1;34m'https://tfhub.dev/tensorflow/bert_en_uncased_preprocess/3'\u001b[0m\u001b[1;33m)\u001b[0m\u001b[1;33m\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n",
|
||||
"\u001b[1;31mModuleNotFoundError\u001b[0m: No module named 'tensorflow_text'"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"import tensorflow_text \n",
|
||||
"import tensorflow_hub as hub\n",
|
||||
"vectorizer = hub.KerasLayer('https://tfhub.dev/tensorflow/bert_en_uncased_preprocess/3')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"{'input_type_ids': <tf.Tensor: shape=(1, 128), dtype=int32, numpy=\n",
|
||||
" array([[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]],\n",
|
||||
" dtype=int32)>,\n",
|
||||
" 'input_word_ids': <tf.Tensor: shape=(1, 128), dtype=int32, numpy=\n",
|
||||
" array([[ 101, 1045, 2293, 19081, 102, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0]], dtype=int32)>,\n",
|
||||
" 'input_mask': <tf.Tensor: shape=(1, 128), dtype=int32, numpy=\n",
|
||||
" array([[1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n",
|
||||
" 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]],\n",
|
||||
" dtype=int32)>}"
|
||||
]
|
||||
},
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"vectorizer(['I love transformers'])"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"使用與原始網路訓練時相同的向量化工具是非常重要的。此外，BERT 向量化工具會返回三個組件：\n",
|
||||
"* `input_word_ids`，這是一個表示輸入句子中標記編號的序列\n",
|
||||
"* `input_mask`，用於顯示序列中哪些部分包含實際輸入，哪些部分是填充。這與 `Masking` 層生成的遮罩類似\n",
|
||||
"* `input_type_ids` 用於語言建模任務，允許在一個序列中指定兩個輸入句子。\n",
|
||||
"\n",
|
||||
"接下來，我們可以實例化 BERT 特徵提取器：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 8,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"bert = hub.KerasLayer('https://tfhub.dev/tensorflow/small_bert/bert_en_uncased_L-4_H-128_A-2/1')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 9,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"pooled_output -> (1, 128)\n",
|
||||
"encoder_outputs -> 4\n",
|
||||
"sequence_output -> (1, 128, 128)\n",
|
||||
"default -> (1, 128)\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"z = bert(vectorizer(['I love transformers']))\n",
|
||||
"for i,x in z.items():\n",
|
||||
" print(f\"{i} -> { len(x) if isinstance(x, list) else x.shape }\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"所以，BERT 層會返回一些有用的結果：\n",
|
||||
"* `pooled_output` 是通過平均序列中所有 token 的結果。你可以將其視為整個網絡的智能語義嵌入。它相當於我們之前模型中的 `GlobalAveragePooling1D` 層的輸出。\n",
|
||||
"* `sequence_output` 是最後一層 transformer 的輸出（對應於我們上面模型中的 `TransformerBlock` 的輸出）。\n",
|
||||
"* `encoder_outputs` 是所有 transformer 層的輸出。由於我們載入的是 4 層的 BERT 模型（從名稱中包含 `4_H` 可以推測出來），它有 4 個張量。最後一個張量與 `sequence_output` 相同。\n",
|
||||
"\n",
|
||||
"現在我們將定義端到端的分類模型。我們將使用*函數式模型定義*，在定義模型輸入後，提供一系列表達式來計算其輸出。我們還會將 BERT 模型的權重設置為不可訓練，只訓練最終的分類器：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 10,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Model: \"model\"\n",
|
||||
"__________________________________________________________________________________________________\n",
|
||||
"Layer (type) Output Shape Param # Connected to \n",
|
||||
"==================================================================================================\n",
|
||||
"input_1 (InputLayer) [(None,)] 0 \n",
|
||||
"__________________________________________________________________________________________________\n",
|
||||
"keras_layer (KerasLayer) {'input_type_ids': ( 0 input_1[0][0] \n",
|
||||
"__________________________________________________________________________________________________\n",
|
||||
"keras_layer_1 (KerasLayer) {'pooled_output': (N 4782465 keras_layer[0][0] \n",
|
||||
" keras_layer[0][1] \n",
|
||||
" keras_layer[0][2] \n",
|
||||
"__________________________________________________________________________________________________\n",
|
||||
"dropout_4 (Dropout) (None, 128) 0 keras_layer_1[0][5] \n",
|
||||
"__________________________________________________________________________________________________\n",
|
||||
"dense_4 (Dense) (None, 4) 516 dropout_4[0][0] \n",
|
||||
"==================================================================================================\n",
|
||||
"Total params: 4,782,981\n",
|
||||
"Trainable params: 516\n",
|
||||
"Non-trainable params: 4,782,465\n",
|
||||
"__________________________________________________________________________________________________\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"inp = keras.Input(shape=(),dtype=tf.string)\n",
|
||||
"x = vectorizer(inp)\n",
|
||||
"x = bert(x)\n",
|
||||
"x = keras.layers.Dropout(0.1)(x['pooled_output'])\n",
|
||||
"out = keras.layers.Dense(4,activation='softmax')(x)\n",
|
||||
"model = keras.models.Model(inp,out)\n",
|
||||
"bert.trainable = False\n",
|
||||
"model.summary()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 11,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"938/938 [==============================] - 528s 559ms/step - loss: 0.8056 - acc: 0.6983 - val_loss: 0.5953 - val_acc: 0.7888\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<tensorflow.python.keras.callbacks.History at 0x7f9bb1e36d00>"
|
||||
]
|
||||
},
|
||||
"execution_count": 11,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model.compile(loss='sparse_categorical_crossentropy',metrics=['acc'], optimizer='adam')\n",
|
||||
"model.fit(ds_train.map(tupelize).batch(128),validation_data=ds_test.map(tupelize).batch(128))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"儘管可訓練的參數很少，但這個過程相當緩慢，因為 BERT 特徵提取器的計算量非常大。看起來我們無法達到合理的準確率，可能是因為訓練不足，或者模型參數不足。\n",
|
||||
"\n",
|
||||
"讓我們嘗試解凍 BERT 的權重並對其進行訓練。這需要非常小的學習率，並且需要更謹慎的訓練策略，包括使用 **warmup** 和 **AdamW** 優化器。我們將使用 `tf-models-official` 套件來創建優化器：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 12,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Model: \"model\"\n",
|
||||
"__________________________________________________________________________________________________\n",
|
||||
"Layer (type) Output Shape Param # Connected to \n",
|
||||
"==================================================================================================\n",
|
||||
"input_1 (InputLayer) [(None,)] 0 \n",
|
||||
"__________________________________________________________________________________________________\n",
|
||||
"keras_layer (KerasLayer) {'input_type_ids': ( 0 input_1[0][0] \n",
|
||||
"__________________________________________________________________________________________________\n",
|
||||
"keras_layer_1 (KerasLayer) {'pooled_output': (N 4782465 keras_layer[0][0] \n",
|
||||
" keras_layer[0][1] \n",
|
||||
" keras_layer[0][2] \n",
|
||||
"__________________________________________________________________________________________________\n",
|
||||
"dropout_4 (Dropout) (None, 128) 0 keras_layer_1[0][5] \n",
|
||||
"__________________________________________________________________________________________________\n",
|
||||
"dense_4 (Dense) (None, 4) 516 dropout_4[0][0] \n",
|
||||
"==================================================================================================\n",
|
||||
"Total params: 4,782,981\n",
|
||||
"Trainable params: 4,782,980\n",
|
||||
"Non-trainable params: 1\n",
|
||||
"__________________________________________________________________________________________________\n",
|
||||
"938/938 [==============================] - 629s 664ms/step - loss: 0.6344 - acc: 0.7658 - val_loss: 0.4876 - val_acc: 0.8247\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<tensorflow.python.keras.callbacks.History at 0x7f9bb0bd0070>"
|
||||
]
|
||||
},
|
||||
"execution_count": 12,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"from official.nlp import optimization \n",
|
||||
"bert.trainable=True\n",
|
||||
"model.summary()\n",
|
||||
"epochs = 3\n",
|
||||
"opt = optimization.create_optimizer(\n",
|
||||
" init_lr=3e-5,\n",
|
||||
" num_train_steps=epochs*len(ds_train),\n",
|
||||
" num_warmup_steps=0.1*epochs*len(ds_train),\n",
|
||||
" optimizer_type='adamw')\n",
|
||||
"\n",
|
||||
"model.compile(loss='sparse_categorical_crossentropy',metrics=['acc'], optimizer=opt)\n",
|
||||
"model.fit(ds_train.map(tupelize).batch(128),validation_data=ds_test.map(tupelize).batch(128))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"如您所見，訓練過程進展相當緩慢——但您可能希望嘗試進行幾個訓練週期（5-10次），並比較我們之前使用的方法，看看是否能獲得最佳結果。\n",
|
||||
"\n",
|
||||
"## Huggingface Transformers 庫\n",
|
||||
"\n",
|
||||
"另一種非常常見（且稍微簡單一些）的使用 Transformer 模型的方法是 [HuggingFace 套件](https://github.com/huggingface/)，它為不同的 NLP 任務提供了簡單的構建模塊。該套件同時支持 Tensorflow 和 PyTorch，後者是另一個非常流行的神經網絡框架。\n",
|
||||
"\n",
|
||||
"> **注意**：如果您對了解 Transformers 庫的工作方式不感興趣——可以直接跳到筆記本的末尾，因為您不會看到與我們之前所做的有實質性不同的內容。我們將重複使用不同的庫和更大的模型來訓練 BERT 模型的相同步驟。因此，這個過程涉及一些相當長的訓練時間，您可能只需要瀏覽一下代碼即可。\n",
|
||||
"\n",
|
||||
"讓我們看看如何使用 [Huggingface Transformers](http://huggingface.co) 解決我們的問題。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"首先，我們需要選擇要使用的模型。除了內建模型之外，Huggingface 還擁有一個[線上模型庫](https://huggingface.co/models)，在那裡你可以找到社群提供的更多預訓練模型。所有這些模型都可以通過提供模型名稱來加載和使用。所需的模型二進制文件會自動下載。\n",
|
||||
"\n",
|
||||
"有時候你可能需要加載自己的模型，這種情況下你可以指定包含所有相關文件的目錄，包括分詞器的參數、`config.json` 文件中的模型參數、二進制權重等。\n",
|
||||
"\n",
|
||||
"從模型名稱開始，我們可以實例化模型和分詞器。讓我們先從分詞器開始：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 2,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import transformers\n",
|
||||
"\n",
|
||||
"# To load the model from Internet repository using model name. \n",
|
||||
"# Use this if you are running from your own copy of the notebooks\n",
|
||||
"bert_model = 'bert-base-uncased' \n",
|
||||
"\n",
|
||||
"# To load the model from the directory on disk. Use this for Microsoft Learn module, because we have\n",
|
||||
"# prepared all required files for you.\n",
|
||||
"#bert_model = './bert'\n",
|
||||
"\n",
|
||||
"tokenizer = transformers.BertTokenizer.from_pretrained(bert_model)\n",
|
||||
"\n",
|
||||
"MAX_SEQ_LEN = 128\n",
|
||||
"PAD_INDEX = tokenizer.convert_tokens_to_ids(tokenizer.pad_token)\n",
|
||||
"UNK_INDEX = tokenizer.convert_tokens_to_ids(tokenizer.unk_token)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"`tokenizer` 對象包含 `encode` 函數，可直接用於編碼文本：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"[101, 23435, 12314, 2003, 1037, 2307, 7705, 2005, 17953, 2361, 102]"
|
||||
]
|
||||
},
|
||||
"execution_count": 3,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"tokenizer.encode('Tensorflow is a great framework for NLP')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"我們也可以使用分詞器來編碼序列，使其適合傳遞給模型，例如包括 `token_ids`、`input_mask` 等字段。我們還可以通過提供 `return_tensors='tf'` 參數來指定我們想要 Tensorflow 張量：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"{'input_ids': <tf.Tensor: shape=(1, 5), dtype=int32, numpy=array([[ 101, 7592, 1010, 2045, 102]], dtype=int32)>, 'token_type_ids': <tf.Tensor: shape=(1, 5), dtype=int32, numpy=array([[0, 0, 0, 0, 0]], dtype=int32)>, 'attention_mask': <tf.Tensor: shape=(1, 5), dtype=int32, numpy=array([[1, 1, 1, 1, 1]], dtype=int32)>}"
|
||||
]
|
||||
},
|
||||
"execution_count": 4,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"tokenizer(['Hello, there'],return_tensors='tf')"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"在我們的案例中，我們將使用名為 `bert-base-uncased` 的預訓練 BERT 模型。*Uncased* 表示該模型對大小寫不敏感。\n",
|
||||
"\n",
|
||||
"在訓練模型時，我們需要提供已分詞的序列作為輸入，因此我們將設計數據處理管道。由於 `tokenizer.encode` 是一個 Python 函數，我們將採用與上一單元相同的方法，使用 `py_function` 來調用它：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 31,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"def process(x):\n",
|
||||
" return tokenizer.encode(x.numpy().decode('utf-8'),return_tensors='tf',padding='max_length',max_length=MAX_SEQ_LEN,truncation=True)[0]\n",
|
||||
"\n",
|
||||
"def process_fn(x):\n",
|
||||
" s = x['title']+' '+x['description']\n",
|
||||
" e = tf.py_function(process,inp=[s],Tout=(tf.int32))\n",
|
||||
" e.set_shape(MAX_SEQ_LEN)\n",
|
||||
" return e,x['label']"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在我們可以使用 `BertForSequenceClassfication` 套件加載實際模型。這確保了我們的模型已經具備分類所需的架構，包括最終的分類器。您會看到一條警告訊息，指出最終分類器的權重尚未初始化，並且模型需要進行預訓練——這完全沒問題，因為這正是我們即將要做的！\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 32,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"model = transformers.TFBertForSequenceClassification.from_pretrained(bert_model,num_labels=4,output_attentions=False)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 33,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Model: \"tf_bert_for_sequence_classification_1\"\n",
|
||||
"_________________________________________________________________\n",
|
||||
"Layer (type) Output Shape Param # \n",
|
||||
"=================================================================\n",
|
||||
"bert (TFBertMainLayer) multiple 109482240 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"dropout_75 (Dropout) multiple 0 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"classifier (Dense) multiple 3076 \n",
|
||||
"=================================================================\n",
|
||||
"Total params: 109,485,316\n",
|
||||
"Trainable params: 109,485,316\n",
|
||||
"Non-trainable params: 0\n",
|
||||
"_________________________________________________________________\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model.summary()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"從 `summary()` 中可以看到，該模型包含了將近 1.1 億個參數！假設我們希望在相對較小的數據集上進行簡單的分類任務，我們可能不希望訓練 BERT 基層：\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 34,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Model: \"tf_bert_for_sequence_classification_1\"\n",
|
||||
"_________________________________________________________________\n",
|
||||
"Layer (type) Output Shape Param # \n",
|
||||
"=================================================================\n",
|
||||
"bert (TFBertMainLayer) multiple 109482240 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"dropout_75 (Dropout) multiple 0 \n",
|
||||
"_________________________________________________________________\n",
|
||||
"classifier (Dense) multiple 3076 \n",
|
||||
"=================================================================\n",
|
||||
"Total params: 109,485,316\n",
|
||||
"Trainable params: 3,076\n",
|
||||
"Non-trainable params: 109,482,240\n",
|
||||
"_________________________________________________________________\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model.layers[0].trainable = False\n",
|
||||
"model.summary()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"現在我們準備開始訓練！\n",
|
||||
"\n",
|
||||
"> **注意**：訓練完整規模的 BERT 模型可能會非常耗時！因此，我們只會訓練前 32 個批次。這只是為了展示模型訓練的設置方式。如果您有興趣嘗試完整規模的訓練，只需移除 `steps_per_epoch` 和 `validation_steps` 參數，並準備耐心等待！\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 30,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"32/32 [==============================] - 142s 4s/step - loss: 1.3896 - acc: 0.2500 - val_loss: 1.3863 - val_acc: 0.2480\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<tensorflow.python.keras.callbacks.History at 0x7f1d40a4b6a0>"
|
||||
]
|
||||
},
|
||||
"execution_count": 30,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"model.compile('adam','sparse_categorical_crossentropy',['acc'])\n",
|
||||
"tf.get_logger().setLevel('ERROR')\n",
|
||||
"model.fit(ds_train.map(process_fn).batch(32),validation_data=ds_test.map(process_fn).batch(32),steps_per_epoch=32,validation_steps=2)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"如果你增加迭代次數並耐心等待，並且訓練多個周期，你可以期待 BERT 分類能夠提供最佳的準確率！這是因為 BERT 已經相當了解語言的結構，我們只需要微調最終的分類器。然而，由於 BERT 是一個大型模型，整個訓練過程需要很長時間，並且需要強大的計算能力！（GPU，最好是多個）。\n",
|
||||
"\n",
|
||||
"> **Note:** 在我們的範例中，我們使用的是最小的預訓練 BERT 模型之一。還有更大的模型可能會產生更好的結果。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## 重點\n",
|
||||
"\n",
|
||||
"在本單元中，我們探討了基於**transformers**的最新模型架構。我們將其應用於文本分類任務，但同樣地，BERT模型也可以用於實體抽取、問答系統以及其他自然語言處理任務。\n",
|
||||
"\n",
|
||||
"Transformer模型代表了自然語言處理領域的最新技術，在大多數情況下，當實現自定義自然語言處理解決方案時，這應該是您首先嘗試的解決方案。然而，如果您希望構建更高級的神經網絡模型，理解本模組中討論的循環神經網絡的基本原理是非常重要的。\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": []
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"\n---\n\n**免責聲明**： \n本文件使用 AI 翻譯服務 [Co-op Translator](https://github.com/Azure/co-op-translator) 進行翻譯。我們致力於提供準確的翻譯，但請注意，自動翻譯可能包含錯誤或不準確之處。應以原始語言的文件作為權威來源。對於關鍵資訊，建議尋求專業人工翻譯。我們對於因使用此翻譯而產生的任何誤解或錯誤解讀概不負責。\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"interpreter": {
|
||||
"hash": "0cb620c6d4b9f7a635928804c26cf22403d89d98d79684e4529119355ee6d5a5"
|
||||
},
|
||||
"kernelspec": {
|
||||
"display_name": "py38_tensorflow",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.8.12"
|
||||
},
|
||||
"coopTranslator": {
|
||||
"original_hash": "ab59c532409774988ab875f2260e8e53",
|
||||
"translation_date": "2025-08-31T10:41:30+00:00",
|
||||
"source_file": "lessons/5-NLP/18-Transformers/TransformersTF.ipynb",
|
||||
"language_code": "tw"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 4
|
||||
}
|
||||
Some files were not shown because too many files have changed in this diff Show More
Loading…
Reference in New Issue