DEVREADME.md

Tips/Information on basic functionality of codebase

one hot labels that appear in losses

one_hot_labels is 2*k wide, as in /gan/tensorflow_gan/python/estimator/gan_estimator.py:_make_gan_model and /gan/tensorflow_gan/python/estimator/tpu_gan_estimator.py:_make_gan_model_fns and more importantly in /gan/tensorflow_gan/python/train.py:acgan_model

And k = num classes, except for a kplusonegan, k = num_classes + 1

to add a new loss

• write fn in /gan/tensorflow_gan/python/losses/losses_impl.py, add to __all__ there
• add in /gan/tensorflow_gan/python/losses/tuple_losses.py like so:

kplusonegan_nll_discriminator_loss = args_to_gan_model(
    tfgan_losses.kplusonegan_nll_discriminator_loss)

If its a --generator_loss_fn.

• add it to the --generator_loss_fn enums in /gan/tensorflow_gan/examples/self_attention_estimator/train_experiment_main.py
• add it to gen_losses hash in get_tpu_estimator and get_gpu_estimator in /gan/tensorflow_gan/examples/self_attention_estimator/estimator_lib.py

Otherwise:

• add a flag in /gan/tensorflow_gan/examples/self_attention_estimator/train_experiment_main.py
• add in /gan/tensorflow_gan/python/train.py in blocks related to:
  • args of gan_loss fn
  • validate args
  • verify config
  • using the loss

Flags info

for K+1 gan, need --critic_type=kplusone_fm so that network will return (features, K+1 logits) tuple

by setting --generator_loss_fn, main loss for discriminator becomes nothing, so add in the other loss through auxiliary