Hands-On Generative Adversarial Networks with Keras

Training loop

To conclude, we define the training loop itself:

    for i in range(n_iterations):        # we are performing D updates only, fix G weights        D.trainable = True        G.trainable = False        for j in range(N_CRITIC_ITERS):            z = np.random.normal(0, 1, size=(batch_size, z_dim))            real_batch = next(data_iterator)            losses_d = D_model.train_on_batch(                [real_batch, z], [minus_ones, ones, dummy])    # we are performing G updates only, fix D weights    D.trainable = False    G.trainable = True    z = np.random.normal(0, 1, size=(batch_size, z_dim))    loss_g = G_model.train_on_batch(z, minus_ones)    if (i % iters_per_checkpoint) == 0:        G.trainable = False        fake_text = G.predict(z_fixed)        log_text(sample(fake_text, id_to_token, word_join), 'fake', i, logger) ...