Hands-On Transfer Learning with Python

Decoder

The final stage of the network is a decoder. In the first two sections of the network, we utilized an encoder and a pretrained model to learn different features and generate an embedding. The output from the fusion layer is a tensor of size H/8 x W/8 x 256, where H and W are the original height and width of the grayscale image (in our case, it was 256 x 256). This input is passed through an eight-layer decoder built using five convolutional and three upsampling layers. The upsampling layers help us to double the size of image using a basic nearest-neighbor approach. The following snippet showcases the decoder section of the network:

#Decoderdec_output = Conv2D(128, (3,3),                        activation='relu',                        padding='same')(fusion_layer_output)dec_output ...