def discriminator(images_and_lbls, unused_conditioning, mode):
"""TF-GAN compatible discriminator."""
del unused_conditioning, mode
images, labels = images_and_lbls['images'], images_and_lbls['labels']
if hparams.debug_params.fake_nets:
# Need discriminator variables and to depend on the generator.
logits = tf.zeros(
[tf.shape(input=images)[0], 20]) * tf.compat.v1.get_variable(
'dummy_d',
initializer=2.0) * tf.reduce_mean(input_tensor=images)
discriminator_vars = ()
else:
num_trainable_variables = len(tf.compat.v1.trainable_variables())
logits, discriminator_vars = dis_module.discriminator(
images, labels, hparams.df_dim, hparams.num_classes)
if num_trainable_variables != len(
tf.compat.v1.trainable_variables()):
# Log the generated variables only in the first time the function is
# called and new variables are generated (it is called twice: once for
# the generated data and once for the real data).
eval_lib.log_and_summarize_variables(
discriminator_vars, 'dvars',
hparams.tpu_params.use_tpu_estimator)
logits.shape.assert_is_compatible_with([None, None])
return logits
def generator(noise, mode):
"""TF-GAN compatible generator function."""
batch_size = tf.shape(input=noise)[0]
is_train = (mode == tf.estimator.ModeKeys.TRAIN)
# Some label trickery.
gen_class_logits = tf.zeros((batch_size, hparams.num_classes))
gen_class_ints = tf.random.categorical(logits=gen_class_logits,
num_samples=1)
gen_sparse_class = tf.squeeze(gen_class_ints, -1)
gen_sparse_class.shape.assert_is_compatible_with([None])
if hparams.debug_params.fake_nets:
gen_imgs = tf.zeros([batch_size, 128, 128, 3]) * tf.get_variable(
'dummy_g', initializer=2.0)
generator_vars = ()
else:
gen_imgs, generator_vars = gen_module.generator(
noise,
gen_sparse_class,
hparams.gf_dim,
hparams.num_classes,
training=is_train)
# Print debug statistics and log the generated variables.
gen_imgs, gen_sparse_class = eval_lib.print_debug_statistics(
gen_imgs, gen_sparse_class, 'generator',
hparams.tpu_params.use_tpu_estimator)
eval_lib.log_and_summarize_variables(
generator_vars, 'gvars', hparams.tpu_params.use_tpu_estimator)
gen_imgs.shape.assert_is_compatible_with([None, 128, 128, 3])
if mode == tf.estimator.ModeKeys.PREDICT:
return gen_imgs
else:
return {'images': gen_imgs, 'labels': gen_sparse_class}
def generator(noise_and_lbls, mode):
"""TF-GAN compatible generator function."""
noise, labs = noise_and_lbls['z'], noise_and_lbls['labels']
batch_size = tf.shape(input=noise)[0]
is_train = (mode == tf.estimator.ModeKeys.TRAIN)
# labs.shape.assert_is_compatible_with([None]) # not correct for gen_images
if hparams.debug_params.fake_nets:
gen_imgs = tf.zeros([
batch_size, flags.FLAGS.image_size, flags.FLAGS.image_size, 3
]) * tf.compat.v1.get_variable('dummy_g', initializer=2.0)
generator_vars = ()
else:
gen_imgs, generator_vars = gen_module.generator(
noise,
labs,
hparams.gf_dim,
hparams.num_classes,
training=is_train)
# Print debug statistics and log the generated variables.
gen_imgs, gen_sparse_class = eval_lib.print_debug_statistics(
gen_imgs, labs, 'generator', hparams.tpu_params.use_tpu_estimator)
eval_lib.log_and_summarize_variables(
generator_vars, 'gvars', hparams.tpu_params.use_tpu_estimator)
gen_imgs.shape.assert_is_compatible_with(
[None, flags.FLAGS.image_size, flags.FLAGS.image_size, 3])
if mode == tf.estimator.ModeKeys.PREDICT and not flags.FLAGS.gen_images_with_margins:
return gen_imgs
else:
return {'images': gen_imgs, 'labels': labs}