Deep Learning for Computer Vision

Training on bottleneck features

The models that are covered in the previous sections are simple and hence, may yield less accuracy. Complex models should be built from them. They cannot be built from scratch. Hence, bottleneck features are extracted and the classifier is trained on them. Bottleneck features are the features that are produced by complex architectures training several million images. The images are done with a forward pass and the pre-final layer features are stored. From these, a simple logistic classifier is trained for classification. Extract the bottleneck layers as follows:

generator = tf.keras.preprocessing.image.ImageDataGenerator(rescale=1. / 255) model = tf.keras.applications.VGG16(include_top=False) train_images = ...