batch_size = 128
updates_per_epoch = 100
max_epoch = 50
exp_name = "mnist_%s" % timestamp
log_dir = os.path.join(root_log_dir, exp_name)
checkpoint_dir = os.path.join(root_checkpoint_dir, exp_name)
mkdir_p(log_dir)
mkdir_p(checkpoint_dir)
dataset = MnistDataset()
latent_spec = [
(Uniform(62), False),
(Categorical(10), True),
(Uniform(1, fix_std=True), True),
(Uniform(1, fix_std=True), True),
]
model = RegularizedGAN(
output_dist=MeanBernoulli(dataset.image_dim),
latent_spec=latent_spec,
batch_size=batch_size,
image_shape=dataset.image_shape,
network_type="mnist",
)
algo = InfoGANTrainer(
model=model,
output_dist = None
network_type = 'dcgan'
if FLAGS.train_dataset == "mnist":
dataset = datasets.MnistDataset()
output_dist = MeanBernoulli(dataset.image_dim)
network_type = 'mnist'
dataset.batch_idx = 100
else:
dataset = datasets.Dataset(name=FLAGS.train_dataset,
batch_size=batch_size,
output_size=FLAGS.output_size)
val_dataset = datasets.Dataset(name=FLAGS.val_dataset,
batch_size=batch_size,
output_size=FLAGS.output_size)
latent_spec = [(Uniform(100), False)]
if FLAGS.categories is not None:
latent_spec.append((Categorical(FLAGS.categories), True))
is_reg = False
for x, y in latent_spec:
if y:
is_reg = True
model = RegularizedGAN(
output_dist=output_dist,
latent_spec=latent_spec,
is_reg=is_reg,
batch_size=batch_size,
image_shape=dataset.image_shape,
network_type=network_type,
def train(configPath):
with open(configPath, 'r') as f:
d = json.load(f)
train_dataset = d['train_dataset']
train_Folder = d['train_Folder']
val_dataset = d['val_dataset']
output_size = d['output_size']
categories = d['categories']
batch_size = d['batch_size']
if d.has_key('improvedGan'):
improvedGan = d['improvedGan']
else:
improvedGan = False
print("Using improvedGAN ", improvedGan)
if d.has_key('exp_name'):
exp_name = d['exp_name']
else:
filename = os.path.split(configPath)[-1]
assert (len(filename.split(".")) == 2)
exp_name = filename.split(".")[0]
arch = d['arch']
if d.has_key('semiSup'):
semiSup = d['semiSup']
else:
semiSup = False
if d.has_key('trainSplit'):
trainSplit = d['trainSplit']
else:
trainSplit = 0.7
now = datetime.datetime.now(dateutil.tz.tzlocal())
timestamp = now.strftime('%Y_%m_%d_%H_%M_%S')
root_log_dir = "logs/" + train_dataset
root_checkpoint_dir = "ckt/" + train_dataset
root_samples_dir = "samples/" + train_dataset
updates_per_epoch = 100
max_epoch = 50
if exp_name is None:
exp_name = "t-%s_v-%s_o-%d" % (train_dataset, val_dataset, output_size)
if not (categories is None):
exp_name = exp_name + "_c-%d" % (categories)
exp_name = exp_name + "_%s" % (timestamp)
print("Experiment Name: %s" % (exp_name))
log_dir = os.path.join(root_log_dir, exp_name)
checkpoint_dir = os.path.join(root_checkpoint_dir, exp_name)
samples_dir = os.path.join(root_samples_dir, exp_name)
mkdir_p(log_dir)
mkdir_p(checkpoint_dir)
mkdir_p(samples_dir)
output_dist = None
network_type = arch
if train_dataset == "mnist":
print("Creating train dataset ")
dataset = datasets.MnistDataset()
output_dist = MeanBernoulli(dataset.image_dim)
print("CREATED train dataset ")
network_type = 'mnist'
print("Creating VAL dataset ")
val_dataset = dataset
elif train_dataset == "dataFolder":
dataset = datasets.DataFolder(train_Folder,
batch_size,
out_size=output_size,
validation_proportion=(1 - trainSplit))
print("Folder datasets created ")
val_dataset = dataset
else:
dataset = datasets.Dataset(name=train_dataset,
batch_size=batch_size,
output_size=output_size)
val_dataset = datasets.Dataset(name=val_dataset,
batch_size=batch_size,
output_size=output_size)
latent_spec = [(Uniform(100), False)]
if categories is not None:
latent_spec.append((Categorical(categories), True))
is_reg = False
for x, y in latent_spec:
if y:
is_reg = True
model = RegularizedGAN(output_dist=output_dist,
latent_spec=latent_spec,
is_reg=is_reg,
batch_size=batch_size,
image_shape=dataset.image_shape,
network_type=network_type,
impr=improvedGan)
if (not improvedGan):
algo = InfoGANTrainer(model=model,
dataset=dataset,
val_dataset=val_dataset,
batch_size=batch_size,
isTrain=True,
exp_name=exp_name,
log_dir=log_dir,
checkpoint_dir=checkpoint_dir,
samples_dir=samples_dir,
max_epoch=max_epoch,
info_reg_coeff=1.0,
generator_learning_rate=2e-3,
discriminator_learning_rate=2e-3,
semiSup=semiSup)
else:
algo = ImprovedGAN(model=model,
dataset=dataset,
val_dataset=val_dataset,
batch_size=batch_size,
isTrain=True,
exp_name=exp_name,
log_dir=log_dir,
checkpoint_dir=checkpoint_dir,
samples_dir=samples_dir,
max_epoch=max_epoch,
info_reg_coeff=1.0,
generator_learning_rate=2e-3,
discriminator_learning_rate=2e-3,
semiSup=semiSup)
#gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.03)
#config=tf.ConfigProto(gpu_options=gpu_options)
#device_name = "/gpu:0"
#with tf.device(device_name):
algo.init_opt()
with tf.Session() as sess:
algo.train(sess)
#DATASET NEEDS TO GET THE TRAIN SPLIT RIGHT
#BE CAREFUL OF THIS. NOT TRUE. JUST NEED NEXT BATCH WORKING
#WHAT DOES INVERSE TRANSFORM DO?
#SWEET. ALL OF THIS LOOKS GOOD.
dataset = HeartDataset()
#THESE DEFINITELY NEED TO BE CHANGED
#
# latent_spec = [
# (Uniform(62), False),
# (Categorical(10), True),
# (Uniform(1, fix_std=True), True),
# (Uniform(1, fix_std=True), True),
# ]
#SAME LATENT SPEC AS REPORTED FOR FACES
latent_spec = [(Uniform(128), False), (Uniform(1, fix_std=True), True),
(Uniform(1, fix_std=True), True),
(Uniform(1, fix_std=True), True),
(Uniform(1, fix_std=True), True),
(Uniform(1, fix_std=True), True)]
#EVERYTHING SEEMS FINE. NETWORK ARCHITECTURE SHOULD
#BE CHANGED.
model = RegularizedGAN(
output_dist=MeanBernoulli(dataset.image_dim),
latent_spec=latent_spec,
batch_size=batch_size,
image_shape=dataset.image_shape,
network_type="heart",
)
def train(model_name, learning_params):
timestamp = get_timestamp()
root_log_dir = os.path.join('logs', model_name)
root_checkpoint_dir = os.path.join('ckt', model_name)
experiment_name = '{}_{}'.format(model_name, timestamp)
log_dir = os.path.join(root_log_dir, experiment_name)
checkpoint_dir = os.path.join(root_checkpoint_dir, experiment_name)
make_exists(log_dir)
make_exists(checkpoint_dir)
batch_size = learning_params['batch_size']
updates_per_epoch = learning_params['updates_per_epoch']
max_epoch = learning_params['max_epoch']
trainer = learning_params['trainer']
if model_name == 'mnist_infogan':
output_dataset = MnistDataset()
latent_spec = [
(Uniform(62), False),
(Categorical(10), True),
(Uniform(1, fix_std=True), True),
(Uniform(1, fix_std=True), True),
]
model = MNISTInfoGAN(
batch_size=batch_size,
output_dataset=output_dataset,
output_dist=MeanBernoulli(output_dataset.image_dim),
latent_spec=latent_spec,
)
elif model_name == 'mnist_wasserstein':
output_dataset = MnistDataset()
latent_spec = [
(Uniform(62), False),
]
model = MNISTInfoGAN(
batch_size=batch_size,
output_dataset=output_dataset,
output_dist=MeanBernoulli(output_dataset.image_dim),
final_activation=None,
latent_spec=latent_spec,
)
elif model_name == 'celebA_infogan':
output_dataset = CelebADataset()
latent_spec = [
(Uniform(128), False),
(Categorical(10), True),
(Categorical(10), True),
(Categorical(10), True),
(Categorical(10), True),
(Categorical(10), True),
(Categorical(10), True),
(Categorical(10), True),
(Categorical(10), True),
(Categorical(10), True),
(Categorical(10), True),
]
model = CelebAInfoGAN(
batch_size=batch_size,
output_dataset=output_dataset,
output_dist=MeanGaussian(output_dataset.image_dim, fix_std=True),
latent_spec=latent_spec,
)
elif model_name == 'celebA_wasserstein':
output_dataset = CelebADataset()
latent_spec = [
(Uniform(128), False),
]
model = CelebAInfoGAN(
batch_size=batch_size,
output_dataset=output_dataset,
output_dist=MeanGaussian(output_dataset.image_dim, fix_std=True),
final_activation=None,
latent_spec=latent_spec,
)
elif model_name == 'horse_zebra':
horse_dataset = HorseOrZebraDataset('horse')
zebra_dataset = HorseOrZebraDataset('zebra')
horse2zebra_model = Horse2Zebra_CycleGAN(
input_dataset=horse_dataset,
batch_size=batch_size,
output_dataset=zebra_dataset,
output_dist=MeanGaussian(zebra_dataset.image_dim, fix_std=True),
final_activation=None,
scope_suffix='_horse2zebra',
)
zebra2horse_model = Horse2Zebra_CycleGAN(
input_dataset=zebra_dataset,
batch_size=batch_size,
output_dataset=horse_dataset,
output_dist=MeanGaussian(zebra_dataset.image_dim, fix_std=True),
final_activation=None,
scope_suffix='_zebra2horse',
)
else:
raise ValueError('Invalid model_name: {}'.format(model_name))
if trainer == 'infogan':
d_optim = tf.train.AdamOptimizer(2e-4, beta1=0.5)
g_optim = tf.train.AdamOptimizer(1e-3, beta1=0.5)
loss = GANLoss()
loss_builder = InfoGANLossBuilder(
model=model,
loss=loss,
batch_size=batch_size,
g_optimizer=g_optim,
d_optimizer=d_optim,
)
algo = GANTrainer(
loss_builder=loss_builder,
exp_name=experiment_name,
log_dir=log_dir,
checkpoint_dir=checkpoint_dir,
max_epoch=max_epoch,
updates_per_epoch=updates_per_epoch,
)
elif trainer == 'wasserstein':
d_optim = tf.train.AdamOptimizer(2e-4, beta1=0.5)
g_optim = tf.train.AdamOptimizer(1e-3, beta1=0.5)
loss = WassersteinGANLoss()
loss_builder = GANLossBuilder(
model=model,
loss=loss,
batch_size=batch_size,
g_optimizer=g_optim,
d_optimizer=d_optim,
)
algo = WassersteinGANTrainer(
loss_builder=loss_builder,
exp_name=experiment_name,
log_dir=log_dir,
checkpoint_dir=checkpoint_dir,
max_epoch=max_epoch,
updates_per_epoch=updates_per_epoch,
)
elif trainer == 'test':
d_optim = tf.train.AdamOptimizer(2e-4, beta1=0.5)
g_optim = tf.train.AdamOptimizer(1e-3, beta1=0.5)
loss = GANLoss()
loss_builder = InfoGANLossBuilder(
model=model,
loss=loss,
dataset=output_dataset,
batch_size=batch_size,
discrim_optimizer=d_optim,
generator_optimizer=g_optim,
)
algo = GANTrainer(
loss_builder=loss_builder,
exp_name=experiment_name,
log_dir=log_dir,
checkpoint_dir=checkpoint_dir,
max_epoch=max_epoch,
updates_per_epoch=updates_per_epoch,
)
elif trainer == 'cycle_gan':
d_optim = tf.train.AdamOptimizer(2e-4, beta1=0.5)
g_optim = tf.train.AdamOptimizer(2e-4, beta1=0.5)
loss = LeastSquaresGANLoss()
horse2zebra_loss_builder = GANLossBuilder(
model=horse2zebra_model,
loss=loss,
batch_size=batch_size,
)
zebra2horse_loss_builder = GANLossBuilder(
model=zebra2horse_model,
loss=loss,
batch_size=batch_size,
)
loss_builders = [horse2zebra_loss_builder, zebra2horse_loss_builder]
loss_builder = CycleGANLossBuilder(
loss_builders,
g_optimizer=g_optim,
d_optimizer=d_optim,
)
algo = GANTrainer(
loss_builder=loss_builder,
exp_name=experiment_name,
log_dir=log_dir,
checkpoint_dir=checkpoint_dir,
max_epoch=max_epoch,
updates_per_epoch=updates_per_epoch,
)
else:
raise ValueError('Invalid trainer: {}'.format(trainer))
algo.train()
mkdir_p(log_dir)
mkdir_p(checkpoint_dir)
# load the rectangle data
rectangles = np.array(pickle.load(open(training_file, 'rb')))
# ShapeDataset has a fixed height and widht of 28 pixels and only one
# matrix per image (grey-scale)
dataset = ShapeDataset(rectangles, None)
# the variables used for the generation of shapes
# Uniform variables are only supported with a value of 1
# Possible would be also e.g. Categorical(10), True) for
# a categorical variable with 10 Steps.
latent_spec = [
(Uniform(size_z), False),
(Uniform(1, fix_std=True), True),
(Uniform(1, fix_std=True), True),
]
model = RegularizedGAN(
output_dist=MeanBernoulli(dataset.image_dim),
latent_spec=latent_spec,
batch_size=batch_size,
image_shape=dataset.image_shape,
network_type="mnist",
)
algo = InfoGANTrainer(
model=model,
dataset=dataset,
# In[3]:
root_log_dir = "logs/celebA"
root_checkpoint_dir = "ckt/celebA"
batch_size = 128
updates_per_epoch = 100 # How often to run the logging.
checkpoint_snapshot_interval = 1000 # Save a snapshot of the model every __ updates.
max_epoch = 50
# In[ ]:
# The "C.3 CelebA" input settings:
# "For this task, we use 10 ten-dimensional categorical code and 128 noise variables, resulting in a concatenated dimension of 228.."
c3_celebA_latent_spec = [
(Uniform(128), False), # Noise
(Categorical(10), True),
(Categorical(10), True),
(Categorical(10), True),
(Categorical(10), True),
(Categorical(10), True),
(Categorical(10), True),
(Categorical(10), True),
(Categorical(10), True),
(Categorical(10), True),
(Categorical(10), True),
]
c3_celebA_image_size = 32
# In[5]:
# In[6]:
root_log_dir = "logs/speech_face"
root_checkpoint_dir = "ckt/speech_face"
batch_size = 128
updates_per_epoch = 50 # How often to run the logging.
checkpoint_snapshot_interval = 1000 # Save a snapshot of the model every __ updates.
max_epoch = 200
# In[7]:
# For now, copy the "C.3 CelebA" input settings:
# "For this task, we use 10 ten-dimensional categorical code and 128 noise variables, resulting in a concatenated dimension of 228.."
c3_celebA_latent_spec = [
(Uniform(128), False), # Noise
(Categorical(10), True),
(Categorical(10), True),
(Categorical(10), True),
(Categorical(10), True),
(Categorical(10), True),
(Categorical(10), True),
(Categorical(10), True),
(Categorical(10), True),
(Categorical(10), True),
(Categorical(10), True),
]
c3_celebA_image_size = 32
# For now, copy the "C.4 Faces" input settings:
# "For this task, we use 5 continuous latent codes and 128 noise variables, so the input to the generator has dimension 133."