For Medium post on Dog Breed Identification

features_extractor_pseudocode.py

class FeaturesExtractor:
    """Runs pretrained model without top layers on dataset and saves generated
    bottleneck features onto disk.
    """
    def __init__(self, build_fn, preprocess_fn, source,
                 target_size=(299, 299, 3), batch_size=128):        
        
        self.build_fn = build_fn
        self.preprocess_fn = preprocess_fn
        self.source = source
        self.target_size = target_size
        self.batch_size = batch_size
        
    def __call__(self, folder, filename, pool='avg'):
        model = self.build_fn(weights='imagenet', include_top=False, pooling=pool)
        stream = self.source(
            folder=folder, target_size=self.target_size,
            batch_size=self.batch_size, infinite=False)
        
        batches = []
        with tqdm.tqdm_notebook(total=stream.steps_per_epoch) as bar:
            for x_batch, y_batch in stream:
                x_preprocessed = self.preprocess_fn(x_batch)
                batch = model.predict_on_batch(x_preprocessed)
                batches.append(batch)
                bar.update(1)
                
        all_features = np.vstack(batches)
        np.save(filename, all_features)
        return filename