Hands-On Transfer Learning with Python

Encoder

Convolutional Neural Networks (CNNs) are amazing image classifiers. They do so by extracting location-invariant features. In this process, they tend to distort the input image.

In the case of image colorization, such a distortion would be disastrous. To this end, we use an encoder to transform our input grayscale image of H x W dimensions to H/8 x W/8. The encoder maintains the aspect ratio of the image through different layers by using zero padding. The following snippet shows the described encoder using Keras:

#Encoderenc_input = Input(shape=(DIM, DIM, 1,))enc_output = Conv2D(64, (3,3),               activation='relu',               padding='same', strides=2)(enc_input)enc_output = Conv2D(128, (3,3),               activation='relu',               padding='same')(enc_output)enc_output ...