Simplify

docs/topics/media-pipeline.rst

@@ -774,60 +774,31 @@ To enable your custom media pipeline component you must add its class import pat
 
    ITEM_PIPELINES = {"myproject.pipelines.MyImagesPipeline": 300}
 
-Content-Based Image Filtering Pipeline
+Content-based image filtering pipeline
 --------------------------------------
 
-This example overrides the ``get_images()`` method of Images Pipeline for content-based filtering. 
-With a TensorFlow_ image classifier, we can build up a dataset featuring target labels through a broad crawl.::
+This example overrides ``get_images()`` to use an image classifier (e.g. a
+TensorFlow_ model) to filter images before storing them. ``self.model`` and
+``self.label`` should be initialized in ``__init__``:
+
+.. code-block:: python
 
     import numpy as np
-    from PIL import Image
     from io import BytesIO
-
-    import scrapy
+    from PIL import Image
     from scrapy.pipelines.images import ImagesPipeline, ImageException
 
-    import tensorflow as tf
-    import tensorflow_hub as hub
 
     class ImageClassifierPipeline(ImagesPipeline):
-        def __init__(self, store_uri, download_func=None, settings=None):
-            super().__init__(store_uri, settings=settings, download_func=download_func)
-
-            self.label = settings.getint("IMAGE_LABEL")
-
-            # Initializing model
-            self.image_size = 224
-            self.model = tf.keras.Sequential([
-                tf.keras.layers.Lambda(lambda x: tf.keras.applications.mobilenet.preprocess_input(x)),
-                hub.KerasLayer("https://tfhub.dev/google/imagenet/mobilenet_v2_130_224/classification/4")
-                ])
-            self.model.build([None, self.image_size, self.image_size, 3])
-
         def get_images(self, response, request, info, *, item=None):
             path = self.file_path(request, response=response, info=info, item=item)
             orig_image = Image.open(BytesIO(response.body))
-
-            # Prepare image for inference
-            img = np.expand_dims(np.array(orig_image.resize([self.image_size]*2, Image.NEAREST)), axis=0)
-
-            # Get top predicted label
-            prediction = self.model.predict(img)[0].argsort()[::-1][0]
-
-            if prediction != self.label:
+            img = np.expand_dims(np.array(orig_image.resize([224, 224])), axis=0)
+            if self.model.predict(img)[0].argsort()[-1] != self.label:
                 raise ImageException("Image does not match label")
-
             image, buf = self.convert_image(orig_image)
             yield path, image, buf
 
-Similar to the first example, add its class import path to the
-:setting:`ITEM_PIPELINES` setting. Also, specify the label_ for filtering.::
-
-   ITEM_PIPELINES = {
-       'myproject.pipelines.ImageClassifierPipeline': 300
-   }
-   IMAGE_LABEL = 2  # Look for goldfish
 
 .. _MD5 hash: https://en.wikipedia.org/wiki/MD5
 .. _TensorFlow: https://tensorflow.org
-.. _label: https://storage.googleapis.com/download.tensorflow.org/data/ImageNetLabels.txt