train_images = (train_images - 127.5) / 127.5
# use tf.data.Dataset to create batches and shuffle --> data pipeline to TF model
train_dataset = tf.data.Dataset.from_tensor_slices(train_images)
train_dataset = train_dataset.shuffle(buffer_size=BUFFER_SIZE)
train_dataset = train_dataset.batch(batch_size=BATCH_SIZE)
print("Shape of batches: {}".format(train_dataset.output_shapes))
# ----- MODEL ----- #
g = Generator()
d = Discriminator()
# defun gives 10s per epoch performance boost
g.call = tf.contrib.eager.defun(g.call)
d.call = tf.contrib.eager.defun(d.call)
# Optimizers
d_optimizer = tf.train.AdamOptimizer(learning_rate=D_LEARNING_RATE)
g_optimizer = tf.train.AdamOptimizer(learning_rate=G_LEARNING_RATE)
"""
# Checkpoints
checkpoint_dir = output_dir
checkpoint_prefix = os.path.join(checkpoint_dir, "ckpt")
checkpoint = tf.train.Checkpoint(
g_optimizer=g_optimizer,
d_optimizer=d_optimizer,
g=g,
d=d
)
"""
