Mastering Computer Vision with TensorFlow 2.x

Training

The key features to take into account while training are as follows:

• Activation: Tanh
• Stochastic gradient descent (SGD) with a mini-batch size of 128
• Leaky ReLU: Slope = 0.2
• Adam optimizer with a learning rate of 0.0002
• Momentum term of 0.5: A value of 0.9 causes oscillation

The following diagram shows the loss term of the DCGAN during the training phase:

Training starts when the generator receives a random input and the generator loss is defined as its ability to produce fake output. The discriminator loss is defined as its ability to separate the real output from the fake output. Gradients are used to update the generator and ...