def test_calculate_frechet_distance(self):
mu1, sigma1 = np.ones((16, )), np.ones((16, 16))
mu2, sigma2 = mu1 * 2, sigma1 * 2
score = fid_utils.calculate_frechet_distance(mu1=mu1,
mu2=mu2,
sigma1=sigma1,
sigma2=sigma2)
assert type(score) == np.float64
# Inputs check
bad_mu2, bad_sigma2 = np.ones((15, 15)), np.ones((15, 15))
with pytest.raises(ValueError):
fid_utils.calculate_frechet_distance(mu1=mu1,
mu2=bad_mu2,
sigma1=sigma1,
sigma2=bad_sigma2)
def test_calculate_frechet_distance(self):
mu1, sigma1 = np.ones((16, )), np.ones((16, 16))
mu2, sigma2 = mu1 * 2, sigma1 * 2
score = fid_utils.calculate_frechet_distance(mu1=mu1,
mu2=mu2,
sigma1=sigma1,
sigma2=sigma2)
assert type(score) == np.float64
def fid_score(num_real_samples,
num_fake_samples,
netG,
dataset,
seed=0,
device=None,
batch_size=50,
verbose=True,
stats_file=None,
log_dir='./log'):
"""
Computes FID stats using functions that store images in memory for speed and fidelity.
Fidelity since by storing images in memory, we don't subject the scores to different read/write
implementations of imaging libraries.
Args:
num_real_samples (int): The number of real images to use for FID.
num_fake_samples (int): The number of fake images to use for FID.
netG (Module): Torch Module object representing the generator model.
device (str/torch.device): Device identifier to use for computation.
seed (int): The random seed to use.
dataset (str/Dataset): The name of the dataset to load if known, or a custom Dataset object
batch_size (int): The batch size to feedforward for inference.
verbose (bool): If True, prints progress.
stats_file (str): The statistics file to load from if there is already one.
log_dir (str): Directory where feature statistics can be stored.
Returns:
float: Scalar FID score.
"""
start_time = time.time()
# Check inputs
if device is None:
device = torch.device('cuda:0' if torch.cuda.is_available() else "cpu")
if isinstance(dataset, str):
default_datasets = {
'cifar10',
'cifar100',
'stl10_48',
'imagenet_32',
'imagenet_128',
'celeba_64',
'celeba_128',
'lsun_bedroom',
'fake_data',
}
if dataset not in default_datasets:
raise ValueError('For default datasets, must be one of {}'.format(
default_datasets))
elif issubclass(type(dataset), torch.utils.data.Dataset):
if stats_file is None:
raise ValueError(
"stats_file to save/load from cannot be empty if using a custom dataset."
)
if not stats_file.endswith('.npz'):
stats_file = stats_file + '.npz'
else:
raise ValueError(
'dataset must be either a Dataset object or a string.')
# Make sure the random seeds are fixed
torch.manual_seed(seed)
random.seed(seed)
np.random.seed(seed)
# Setup directories
inception_path = os.path.join(log_dir, 'metrics', 'inception_model')
# Setup the inception graph
inception_utils.create_inception_graph(inception_path)
# Start producing statistics for real and fake images
# if device and device.index is not None:
# # Avoid unbounded memory usage
# gpu_options = tf.compat.v1.GPUOptions(allow_growth=True,
# per_process_gpu_memory_fraction=0.15,
# visible_device_list=str(device.index))
# config = tf.compat.v1.ConfigProto(gpu_options=gpu_options)
# else:
# config = tf.compat.v1.ConfigProto(device_count={'GPU': 0})
config = tf.compat.v1.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.2
config.gpu_options.allow_growth = True
with tf.compat.v1.Session(config=config) as sess:
sess.run(tf.compat.v1.global_variables_initializer())
m_real, s_real = compute_real_dist_stats(num_samples=num_real_samples,
sess=sess,
dataset=dataset,
batch_size=batch_size,
verbose=verbose,
stats_file=stats_file,
log_dir=log_dir,
seed=seed)
m_fake, s_fake = compute_gen_dist_stats(netG=netG,
num_samples=num_fake_samples,
sess=sess,
device=device,
seed=seed,
batch_size=batch_size,
verbose=verbose)
FID_score = fid_utils.calculate_frechet_distance(mu1=m_real,
sigma1=s_real,
mu2=m_fake,
sigma2=s_fake)
print("INFO: FID: {} [Time Taken: {:.4f} secs]".format(
FID_score,
time.time() - start_time))
return float(FID_score)
def fid_score(num_real_samples,
num_fake_samples,
netG,
device,
seed,
dataset_name,
batch_size=50,
verbose=True,
stats_file=None,
log_dir='./log'):
"""
Computes FID stats using functions that store images in memory for speed and fidelity.
Fidelity since by storing images in memory, we don't subject the scores to different read/write
implementations of imaging libraries.
Args:
num_real_samples (int): The number of real images to use for FID.
num_fake_samples (int): The number of fake images to use for FID.
netG (Module): Torch Module object representing the generator model.
device (str): Device identifier to use for computation.
seed (int): The random seed to use.
dataset_name (str): The name of the dataset to load.
batch_size (int): The batch size to feedforward for inference.
verbose (bool): If True, prints progress.
stats_file (str): The statistics file to load from if there is already one.
log_dir (str): Directory where feature statistics can be stored.
Returns:
float: Scalar FID score.
"""
start_time = time.time()
# Make sure the random seeds are fixed
torch.manual_seed(seed)
random.seed(seed)
np.random.seed(seed)
# Setup directories
inception_path = os.path.join(log_dir, 'metrics', 'inception_model')
# Setup the inception graph
inception_utils.create_inception_graph(inception_path)
# Start producing statistics for real and fake images
if device and device.index is not None:
# Avoid unbounded memory usage
gpu_options = tf.GPUOptions(allow_growth=True,
per_process_gpu_memory_fraction=0.15,
visible_device_list=str(device.index))
config = tf.ConfigProto(gpu_options=gpu_options)
else:
config = tf.ConfigProto(device_count={'GPU': 0})
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
m_real, s_real = compute_real_dist_stats(num_samples=num_real_samples,
sess=sess,
dataset_name=dataset_name,
batch_size=batch_size,
verbose=verbose,
stats_file=stats_file,
log_dir=log_dir,
seed=seed)
m_fake, s_fake = compute_gen_dist_stats(netG=netG,
num_samples=num_fake_samples,
sess=sess,
device=device,
seed=seed,
batch_size=batch_size,
verbose=verbose)
FID_score = fid_utils.calculate_frechet_distance(mu1=m_real,
sigma1=s_real,
mu2=m_fake,
sigma2=s_fake)
print("INFO: FID Score: {} [Time Taken: {:.4f} secs]".format(
FID_score,
time.time() - start_time))
return float(FID_score)