MNIST-Keras-GAN-Training.py

EPOCHS       = 50
BATCH_SIZE   = 128
BATCH_BLOCKS = x_train.shape[0] // BATCH_SIZE

discriminator_losses = []
gan_losses           = []

for epoch in range(EPOCHS):
    for batch_index in range(BATCH_BLOCKS):
        batch         = x_train[batch_index * BATCH_SIZE:(batch_index + 1) * BATCH_SIZE].reshape(BATCH_SIZE, 784)
        latent_noise  = np.random.normal(0, 1, size=[BATCH_SIZE, LATENT_SIZE])
        generated     = generator.predict(latent_noise)
        batch_images  = np.concatenate([batch, generated])
        labels        = np.zeros(2 * BATCH_SIZE)
        labels[:BATCH_SIZE] = 0.98
        labels[BATCH_SIZE:] = 0.02
        discriminator.trainable = True
        discriminator_loss = discriminator.train_on_batch(batch_images, labels)
        discriminator.trainable = False
        latent_noise  = np.random.normal(0, 1, size=[BATCH_SIZE, LATENT_SIZE])
        labels        = np.ones(BATCH_SIZE)
        gan_loss = gan.train_on_batch(latent_noise, labels)

    discriminator_losses.append(discriminator_loss)
    gan_losses.append(gan_loss)

    print(f"[{epoch + 1}|{EPOCHS}] Generator loss: {gan_loss:.6f} :: Discriminator loss: {discriminator_loss:.6f}.")
    if epoch % 5 == 0:
        plot_generated_images(epoch, generator, LATENT_SIZE)

plot_losses(epoch, discriminator_losses, gan_losses)