def GetGANGridSearch(dataset_name, training_steps, num_seeds):
"""Standard GAN grid search used in the paper."""
config = { # 48 x num_seeds workers
"gan_type": [consts.GAN_WITH_PENALTY, consts.WGAN_WITH_PENALTY],
"penalty_type": [consts.NO_PENALTY, consts.WGANGP_PENALTY],
"discriminator_normalization": [
consts.NO_NORMALIZATION, consts.SPECTRAL_NORM],
"architecture": consts.DCGAN_ARCH,
"dataset": dataset_name,
"tf_seed": list(range(num_seeds)),
"training_steps": training_steps,
"save_checkpoint_steps": 20000,
"batch_size": 64,
"optimizer": "adam",
"z_dim": 64,
"__initial_trials": json.dumps(task_utils.CrossProduct({
"learning_rate": 0.0001,
"lambda": [1, 10],
("beta1", "beta2", "disc_iters"): [
(0.5, 0.9, 5), (0.5, 0.999, 5), (0.9, 0.999, 5)],
})),
}
return config
def CreateCrossProductAndAddDefaultParams(config):
tasks = task_utils.CrossProduct(config)
# Add GAN and dataset specific hyperparams.
for task in tasks:
defaults = GetDefaultParams(
params.GetParameters(task["gan_type"], "wide"))
defaults.update(task)
task.update(defaults)
return tasks
def RepeatExp(gan_type, num_repeat):
"""Experiment where 1 fixed gan is trained with num_repeat seeds."""
config = {
"gan_type": [gan_type],
"dataset": ["mnist", "fashion-mnist"],
"training_steps": [20 * 60000 // BATCH_SIZE],
"save_checkpoint_steps": [5 * 60000 // BATCH_SIZE],
"batch_size": [BATCH_SIZE],
"z_dim": [64],
"tf_seed": range(num_repeat)
}
tasks = task_utils.CrossProduct(config)
# Add GAN and dataset specific hyperparams.
for task in tasks:
task.update(
GetDefaultParams(
params.GetParameters(task["gan_type"], task["dataset"])))
return tasks
def TestExp():
"""Run for one epoch over all tested GANs."""
config = {
"gan_type": [
"GAN", "GAN_MINMAX", "WGAN", "WGAN_GP", "DRAGAN", "VAE", "LSGAN",
"BEGAN"
],
"dataset": ["mnist"],
"training_steps": [60000 // BATCH_SIZE],
"save_checkpoint_steps": [10000],
"batch_size": [BATCH_SIZE],
"tf_seed": [42],
"z_dim": [64],
}
tasks = task_utils.CrossProduct(config)
# Add GAN and dataset specific hyperparams.
for task in tasks:
task.update(
GetDefaultParams(
params.GetParameters(task["gan_type"], task["dataset"])))
return tasks
def GetMultiGANGridSearch(dataset_name, training_steps, num_seeds, aggregate):
"""Standard MultiGAN grid search used in the paper."""
config = { # 42 x num_seeds workers
"gan_type":
"MultiGAN",
"penalty_type":
consts.WGANGP_PENALTY,
"discriminator_normalization":
[consts.NO_NORMALIZATION, consts.SPECTRAL_NORM],
"architecture":
consts.DCGAN_ARCH,
"dataset":
dataset_name,
"tf_seed":
list(range(num_seeds)),
"training_steps":
training_steps,
"save_checkpoint_steps":
20000,
"batch_size":
64,
"optimizer":
"adam",
# Model params.
"aggregate":
aggregate,
"__initial_trials":
json.dumps(
task_utils.CrossProduct({
("n_blocks", "share_block_weights", "n_heads", "k"): [
(0, False, 0, 3),
(0, False, 0, 4), # M-GAN [3, 4]
(0, False, 0, 5), # M-GAN 5
(1, False, 1, 3),
(1, False, 2, 3), # RM-GAN 3
(2, False, 1, 3),
(2, True, 1, 3), # RM-GAN 3
(2, False, 2, 3),
(2, True, 2, 3), # RM-GAN 3
(1, False, 1, 4),
(1, False, 2, 4), # RM-GAN 4
(2, False, 1, 4),
(2, True, 1, 4), # RM-GAN 4
(2, False, 2, 4),
(2, True, 2, 4), # RM-GAN 4
(1, False, 1, 5),
(1, False, 2, 5), # RM-GAN 5
(2, False, 1, 5),
(2, True, 1, 5), # RM-GAN 5
(2, False, 2, 5),
(2, True, 2, 5), # RM-GAN 5
],
("z_dim", "embedding_dim"): [(64, 32)],
"learning_rate":
0.0001,
"lambda":
1,
("beta1", "beta2", "disc_iters"): [(0.9, 0.999, 5)],
}))
}
return config
def GetMetaTasks(experiment_name):
"""Returns meta options to be used for study generation.
Args:
experiment_name: name of an experiment
Raises:
ValueError: When experiment is not found.
"""
if experiment_name == "multi_gan-debug":
meta_config = {
"gan_type": [
"MultiGAN",
],
"penalty_type": [consts.WGANGP_PENALTY],
"discriminator_normalization": [consts.SPECTRAL_NORM],
"architecture": consts.DCGAN_ARCH,
"dataset": ["multi-mnist-3-uniform"],
"sampler": [
"rs",
],
"tf_seed": [0],
"training_steps": [1000],
"save_checkpoint_steps": [100],
"batch_size": [64],
"optimizer": ["adam"],
"learning_rate": [0.0001],
"lambda": [10],
"beta1": [0.5],
"beta2": [0.9],
"disc_iters": [5],
# Model params.
"k": [3],
"aggregate": ["sum_clip"],
"n_heads": 4,
"n_blocks": 2,
"share_block_weights": True,
"embedding_dim": 32,
}
#######################
# PAPER EXPERIMENTS #
#######################
# MULTI-MNIST
# 240
elif experiment_name == "gan-base-experiment-paper":
meta_config = GetGANGridSearch(dataset_name="multi-mnist-3-uniform",
training_steps=1000000,
num_seeds=5)
# 210
elif experiment_name == "multi_gan-base-experiment-paper":
meta_config = GetMultiGANGridSearch(
dataset_name="multi-mnist-3-uniform",
training_steps=1000000,
num_seeds=5,
aggregate="sum_clip")
# 480
elif experiment_name == "gan-relational-experiment-paper":
meta_config = GetGANGridSearch(dataset_name=[
"multi-mnist-3-triplet", "multi-mnist-3-uniform-rgb-occluded"
],
training_steps=1000000,
num_seeds=5)
# 210
elif experiment_name == "multi_gan-relational-experiment-triplet-paper":
meta_config = GetMultiGANGridSearch(
dataset_name="multi-mnist-3-triplet",
training_steps=1000000,
num_seeds=5,
aggregate="sum_clip")
# 210
elif experiment_name == "multi_gan-relational-experiment-rgb-occluded-paper":
meta_config = GetMultiGANGridSearch(
dataset_name="multi-mnist-3-uniform-rgb-occluded",
training_steps=1000000,
num_seeds=5,
aggregate="implicit_alpha")
# CIFAR 10
# 240
elif experiment_name == "gan-background-experiment-paper":
meta_config = GetGANGridSearch(
dataset_name="multi-mnist-3-uniform-rgb-occluded-cifar10",
training_steps=1000000,
num_seeds=5)
# 210
elif experiment_name == "multi_gan-background-experiment-paper":
meta_config = GetMultiGANGridSearchKPlusOne(
dataset_name="multi-mnist-3-uniform-rgb-occluded-cifar10",
training_steps=1000000,
num_seeds=5,
aggregate="alpha")
# 420
elif experiment_name == "multi_gan_bg-background-experiment-paper":
meta_config = GetMultiGANBackgroundGridSearch(
dataset_name="multi-mnist-3-uniform-rgb-occluded-cifar10",
training_steps=1000000,
num_seeds=5,
aggregate="alpha")
# CLEVR
# 240
elif experiment_name == "gan-clevr-experiment-paper":
meta_config = GetGANGridSearch(dataset_name="clevr",
training_steps=1000000,
num_seeds=5)
# 210
elif experiment_name == "multi_gan-clevr-experiment-paper":
meta_config = GetMultiGANGridSearchKPlusOne(dataset_name="clevr",
training_steps=1000000,
num_seeds=5,
aggregate="alpha")
# 420
elif experiment_name == "multi_gan_bg-clevr-experiment-paper":
meta_config = GetMultiGANBackgroundGridSearch(dataset_name="clevr",
training_steps=1000000,
num_seeds=5,
aggregate="alpha")
# 200
elif experiment_name == "multi_gan_bg-clevr-rs200k-paper":
meta_config = GetMultiGANBackgroundRandomSearch(dataset_name="clevr",
training_steps=200000,
aggregate="alpha",
num_tasks=200)
else:
raise ValueError("Unknown study-based experiment %s." %
experiment_name)
options = task_utils.CrossProduct(meta_config)
return options
