Hands-On Generative Adversarial Networks with Keras

Training loops

Unlike traditional GAN setups, the training loop for the Progressive Growing of GANs methodology requires transitioning from a block of lower resolution to a block of a higher resolution. Given that blocks at lower resolution consume less memory, we can also modify the batch size according to the current block at hand:

    while cur_nimg < total_kimg * 1000:        # block processing        kimg = cur_nimg / 1000.0        phase_idx = int(np.floor((kimg + transition_kimg) / phase_kdur))        phase_idx = max(phase_idx, 0.0)        phase_kimg = phase_idx * phase_kdur        # update batch size and ones vector if we switched phases        if phase_idx_prev < phase_idx:            batch_size = MINIBATCH_OVERWRITES[phase_idx]            train_iterator = iterate_minibatches(glob_str, batch_size) ones ...