def main():
# Load a batch of images (to feed to the discriminator)
dataset_iterator = load_image_batch(args.img_dir, batch_size=args.batch_size, n_threads=args.num_data_threads)
# Initialize generator and discriminator models
generator = Generator_Model()
discriminator = Discriminator_Model()
# For saving/loading models
checkpoint_dir = './checkpoints'
checkpoint_prefix = os.path.join(checkpoint_dir, "ckpt")
checkpoint = tf.train.Checkpoint(generator=generator, discriminator=discriminator)
manager = tf.train.CheckpointManager(checkpoint, checkpoint_dir, max_to_keep=3)
# Ensure the output directory exists
if not os.path.exists(args.out_dir):
os.makedirs(args.out_dir)
if args.restore_checkpoint or args.mode == 'test':
checkpoint.restore(manager.latest_checkpoint)
try:
# Specify an invalid GPU device
with tf.device('/device:' + args.device):
if args.mode == 'train':
for epoch in range(0, args.num_epochs):
print('========================== EPOCH %d ==========================' % epoch)
avg_fid = train(generator, discriminator, dataset_iterator, manager)
print("Average FID for Epoch: " + str(avg_fid))
# Save at the end of the epoch, too
print("**** SAVING CHECKPOINT AT END OF EPOCH ****")
manager.save()
if args.mode == 'test':
test(generator)
except RuntimeError as e:
print(e)
with tf.GradientTape() as tape:
means, eps, outputs = wae.call(batch)
loss = wae.loss(means, eps, batch, outputs)
gradients = tape.gradient(loss, wae.trainable_variables)
wae.optimizer.apply_gradients(zip(gradients, wae.trainable_variables))
print(iteration, 'loss', loss)
#save model
# if iteration % 1000 == 0:
# with open('loss_0505/loss_wae_'+str(wae.lamb)+'_'+str(epoch)+'_'+str(iteration), 'wb') as fp:
# pickle.dump(all_loss,fp)
# wae.save_weights(filepath = 'trained_models/wae_'+str(wae.lamb)+'_'+str(epoch)+'_'+str(iteration)+'.h5')
# print('**** LOSS: %g ****' % loss)
#load in data
dataset_iterator = load_image_batch('celebA', batch_size=100, n_threads=2)
#defime z_dim
z_dim = 64
#batch size
batch_size = 100
u = tf.convert_to_tensor(np.zeros(z_dim), dtype='float32')
z_dim = 64
temp = np.zeros((z_dim, z_dim))
for i in range(z_dim):
temp[i, i] = 1.0
v = tf.convert_to_tensor(temp, dtype='float32')
#lamb = 0.001WW
wae = WAE(z_dim, u, v, lamb)
def main():
# Load images
rock = load_image_batch(args.img_dir + '/rock')
rap = load_image_batch(args.img_dir + '/rap')
jazz = load_image_batch(args.img_dir + '/jazz')
# generate labels and make full dataset
genre_labels = np.zeros(((len(rock) + len(rap) + len(jazz)), ))
dataset = np.zeros(((len(rock) + len(rap) + len(jazz)), rock[0].shape[0],
rock[0].shape[1], rock[0].shape[2]))
# concatenating is super slow, so we do this
for i in range(len(rock)):
dataset[i] = rock[i]
offset = len(rock)
for i in range(len(rap)):
genre_labels[i + offset] = 1
dataset[i + offset] = rap[i]
offset = len(rock) + len(jazz)
for i in range(len(jazz)):
genre_labels[i + offset] = 2
dataset[i + offset] = jazz[i]
dataset = tf.convert_to_tensor(dataset)
genre_labels = tf.convert_to_tensor(genre_labels)
# Initialize models
generator = Generator_Model()
discriminator = Discriminator_Model()
mapping_net = Mapping_Model()
noise_net = make_noise_scale_net()
adain_net = ADAin_Model()
# For saving/loading models
checkpoint_dir = './checkpoints'
checkpoint_prefix = os.path.join(checkpoint_dir, "ckpt")
checkpoint = tf.train.Checkpoint(generator=generator,
discriminator=discriminator)
manager = tf.train.CheckpointManager(checkpoint,
checkpoint_dir,
max_to_keep=3)
# Ensure the output directory exists
if not os.path.exists(args.out_dir):
os.makedirs(args.out_dir)
if args.restore_checkpoint or args.mode == 'test':
# restores the latest checkpoint using from the manager
checkpoint.restore(manager.latest_checkpoint)
try:
# Specify an invalid GPU device
with tf.device('/device:' + args.device):
if args.mode == 'train':
for epoch in range(0, args.num_epochs):
print(
'========================== EPOCH %d =========================='
% epoch)
avg_fid = train(generator, discriminator, dataset,
genre_labels, manager, mapping_net,
noise_net, adain_net)
print("Average FID for Epoch: " + str(avg_fid))
# Save at the end of the epoch, too
print("**** SAVING CHECKPOINT AT END OF EPOCH ****")
manager.save()
if args.mode == 'test':
test(generator, args.batch_size, args.z_dim, args.out_dir)
except RuntimeError as e:
print(e)