def main():
batch_size, channels, height, width = 100, 12, 64, 64
inputs = torch.rand(batch_size, channels, height, width)
net = ConvAttentionNet(in_channels=channels,
out_channels=2 * channels,
hidden_channels=32,
num_blocks=4)
print(utils.get_num_parameters(net))
outputs = net(inputs)
print(outputs.shape)
def create_flow(c, h, w, flow_checkpoint, _log):
distribution = distributions.StandardNormal((c * h * w, ))
transform = create_transform(c, h, w)
flow = flows.Flow(transform, distribution)
_log.info('There are {} trainable parameters in this model.'.format(
utils.get_num_parameters(flow)))
if flow_checkpoint is not None:
flow.load_state_dict(torch.load(flow_checkpoint))
_log.info('Flow state loaded from {}'.format(flow_checkpoint))
return flow
def main():
batch_size, channels, height, width = 100, 12, 64, 64
inputs = torch.rand(batch_size, channels, height, width)
context = torch.rand(batch_size, channels // 2, height, width)
net = ConvResidualNet(in_channels=channels,
out_channels=2 * channels,
hidden_channels=32,
context_channels=channels // 2,
num_blocks=2,
dropout_probability=0.1,
use_batch_norm=True)
print(utils.get_num_parameters(net))
outputs = net(inputs, context)
print(outputs.shape)
def create_transform():
transform = transforms.CompositeTransform([
transforms.CompositeTransform(
[create_linear_transform(),
create_base_transform(i)]) for i in range(args.num_flow_steps)
] + [create_linear_transform()])
return transform
# create model
distribution = distributions.StandardNormal((features, ))
transform = create_transform()
flow = flows.Flow(transform, distribution).to(device)
n_params = utils.get_num_parameters(flow)
print('There are {} trainable parameters in this model.'.format(n_params))
# create optimizer
optimizer = optim.Adam(flow.parameters(), lr=args.learning_rate)
if args.anneal_learning_rate:
scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer,
args.num_training_steps,
0)
else:
scheduler = None
# create summary writer and write to log directory
timestamp = cutils.get_timestamp()
if cutils.on_cluster():
timestamp += '||{}'.format(os.environ['SLURM_JOB_ID'])
def run(seed):
assert torch.cuda.is_available()
device = torch.device('cuda')
torch.set_default_tensor_type('torch.cuda.FloatTensor')
np.random.seed(seed)
torch.manual_seed(seed)
# Create training data.
data_transform = tvtransforms.Compose(
[tvtransforms.ToTensor(),
tvtransforms.Lambda(torch.bernoulli)])
if args.dataset_name == 'mnist':
dataset = datasets.MNIST(root=os.path.join(utils.get_data_root(),
'mnist'),
train=True,
download=True,
transform=data_transform)
test_dataset = datasets.MNIST(root=os.path.join(
utils.get_data_root(), 'mnist'),
train=False,
download=True,
transform=data_transform)
elif args.dataset_name == 'fashion-mnist':
dataset = datasets.FashionMNIST(root=os.path.join(
utils.get_data_root(), 'fashion-mnist'),
train=True,
download=True,
transform=data_transform)
test_dataset = datasets.FashionMNIST(root=os.path.join(
utils.get_data_root(), 'fashion-mnist'),
train=False,
download=True,
transform=data_transform)
elif args.dataset_name == 'omniglot':
dataset = data_.OmniglotDataset(split='train',
transform=data_transform)
test_dataset = data_.OmniglotDataset(split='test',
transform=data_transform)
elif args.dataset_name == 'emnist':
rotate = partial(tvF.rotate, angle=-90)
hflip = tvF.hflip
data_transform = tvtransforms.Compose([
tvtransforms.Lambda(rotate),
tvtransforms.Lambda(hflip),
tvtransforms.ToTensor(),
tvtransforms.Lambda(torch.bernoulli)
])
dataset = datasets.EMNIST(root=os.path.join(utils.get_data_root(),
'emnist'),
split='letters',
train=True,
transform=data_transform,
download=True)
test_dataset = datasets.EMNIST(root=os.path.join(
utils.get_data_root(), 'emnist'),
split='letters',
train=False,
transform=data_transform,
download=True)
else:
raise ValueError
if args.dataset_name == 'omniglot':
split = -1345
elif args.dataset_name == 'emnist':
split = -20000
else:
split = -10000
indices = np.arange(len(dataset))
np.random.shuffle(indices)
train_indices, val_indices = indices[:split], indices[split:]
train_sampler = SubsetRandomSampler(train_indices)
val_sampler = SubsetRandomSampler(val_indices)
train_loader = data.DataLoader(
dataset=dataset,
batch_size=args.batch_size,
sampler=train_sampler,
num_workers=4 if args.dataset_name == 'emnist' else 0)
train_generator = data_.batch_generator(train_loader)
val_loader = data.DataLoader(dataset=dataset,
batch_size=1024,
sampler=val_sampler,
shuffle=False,
drop_last=False)
val_batch = next(iter(val_loader))[0]
test_loader = data.DataLoader(
test_dataset,
batch_size=16,
shuffle=False,
drop_last=False,
)
# from matplotlib import pyplot as plt
# from experiments import cutils
# from torchvision.utils import make_grid
# fig, ax = plt.subplots(1, 1, figsize=(5, 5))
# cutils.gridimshow(make_grid(val_batch[:64], nrow=8), ax)
# plt.show()
# quit()
def create_linear_transform():
if args.linear_type == 'lu':
return transforms.CompositeTransform([
transforms.RandomPermutation(args.latent_features),
transforms.LULinear(args.latent_features, identity_init=True)
])
elif args.linear_type == 'svd':
return transforms.SVDLinear(args.latent_features,
num_householder=4,
identity_init=True)
elif args.linear_type == 'perm':
return transforms.RandomPermutation(args.latent_features)
else:
raise ValueError
def create_base_transform(i, context_features=None):
if args.prior_type == 'affine-coupling':
return transforms.AffineCouplingTransform(
mask=utils.create_alternating_binary_mask(
features=args.latent_features, even=(i % 2 == 0)),
transform_net_create_fn=lambda in_features, out_features: nn_.
ResidualNet(in_features=in_features,
out_features=out_features,
hidden_features=args.hidden_features,
context_features=context_features,
num_blocks=args.num_transform_blocks,
activation=F.relu,
dropout_probability=args.dropout_probability,
use_batch_norm=args.use_batch_norm))
elif args.prior_type == 'rq-coupling':
return transforms.PiecewiseRationalQuadraticCouplingTransform(
mask=utils.create_alternating_binary_mask(
features=args.latent_features, even=(i % 2 == 0)),
transform_net_create_fn=lambda in_features, out_features: nn_.
ResidualNet(in_features=in_features,
out_features=out_features,
hidden_features=args.hidden_features,
context_features=context_features,
num_blocks=args.num_transform_blocks,
activation=F.relu,
dropout_probability=args.dropout_probability,
use_batch_norm=args.use_batch_norm),
num_bins=args.num_bins,
tails='linear',
tail_bound=args.tail_bound,
apply_unconditional_transform=args.
apply_unconditional_transform,
)
elif args.prior_type == 'affine-autoregressive':
return transforms.MaskedAffineAutoregressiveTransform(
features=args.latent_features,
hidden_features=args.hidden_features,
context_features=context_features,
num_blocks=args.num_transform_blocks,
use_residual_blocks=True,
random_mask=False,
activation=F.relu,
dropout_probability=args.dropout_probability,
use_batch_norm=args.use_batch_norm)
elif args.prior_type == 'rq-autoregressive':
return transforms.MaskedPiecewiseRationalQuadraticAutoregressiveTransform(
features=args.latent_features,
hidden_features=args.hidden_features,
context_features=context_features,
num_bins=args.num_bins,
tails='linear',
tail_bound=args.tail_bound,
num_blocks=args.num_transform_blocks,
use_residual_blocks=True,
random_mask=False,
activation=F.relu,
dropout_probability=args.dropout_probability,
use_batch_norm=args.use_batch_norm)
else:
raise ValueError
# ---------------
# prior
# ---------------
def create_prior():
if args.prior_type == 'standard-normal':
prior = distributions_.StandardNormal((args.latent_features, ))
else:
distribution = distributions_.StandardNormal(
(args.latent_features, ))
transform = transforms.CompositeTransform([
transforms.CompositeTransform(
[create_linear_transform(),
create_base_transform(i)])
for i in range(args.num_flow_steps)
])
transform = transforms.CompositeTransform(
[transform, create_linear_transform()])
prior = flows.Flow(transform, distribution)
return prior
# ---------------
# inputs encoder
# ---------------
def create_inputs_encoder():
if args.approximate_posterior_type == 'diagonal-normal':
inputs_encoder = None
else:
inputs_encoder = nn_.ConvEncoder(
context_features=args.context_features,
channels_multiplier=16,
dropout_probability=args.dropout_probability_encoder_decoder)
return inputs_encoder
# ---------------
# approximate posterior
# ---------------
def create_approximate_posterior():
if args.approximate_posterior_type == 'diagonal-normal':
context_encoder = nn_.ConvEncoder(
context_features=args.context_features,
channels_multiplier=16,
dropout_probability=args.dropout_probability_encoder_decoder)
approximate_posterior = distributions_.ConditionalDiagonalNormal(
shape=[args.latent_features], context_encoder=context_encoder)
else:
context_encoder = nn.Linear(args.context_features,
2 * args.latent_features)
distribution = distributions_.ConditionalDiagonalNormal(
shape=[args.latent_features], context_encoder=context_encoder)
transform = transforms.CompositeTransform([
transforms.CompositeTransform([
create_linear_transform(),
create_base_transform(
i, context_features=args.context_features)
]) for i in range(args.num_flow_steps)
])
transform = transforms.CompositeTransform(
[transform, create_linear_transform()])
approximate_posterior = flows.Flow(
transforms.InverseTransform(transform), distribution)
return approximate_posterior
# ---------------
# likelihood
# ---------------
def create_likelihood():
latent_decoder = nn_.ConvDecoder(
latent_features=args.latent_features,
channels_multiplier=16,
dropout_probability=args.dropout_probability_encoder_decoder)
likelihood = distributions_.ConditionalIndependentBernoulli(
shape=[1, 28, 28], context_encoder=latent_decoder)
return likelihood
prior = create_prior()
approximate_posterior = create_approximate_posterior()
likelihood = create_likelihood()
inputs_encoder = create_inputs_encoder()
model = vae.VariationalAutoencoder(
prior=prior,
approximate_posterior=approximate_posterior,
likelihood=likelihood,
inputs_encoder=inputs_encoder)
# with torch.no_grad():
# # elbo = model.stochastic_elbo(val_batch[:16].to(device)).mean()
# # print(elbo)
# elbo = model.stochastic_elbo(val_batch[:16].to(device), num_samples=100).mean()
# print(elbo)
# log_prob = model.log_prob_lower_bound(val_batch[:16].to(device), num_samples=1200).mean()
# print(log_prob)
# quit()
n_params = utils.get_num_parameters(model)
print('There are {} trainable parameters in this model.'.format(n_params))
optimizer = optim.Adam(model.parameters(), lr=args.learning_rate)
scheduler = optim.lr_scheduler.CosineAnnealingLR(
optimizer=optimizer, T_max=args.num_training_steps, eta_min=0)
def get_kl_multiplier(step):
if args.kl_multiplier_schedule == 'constant':
return args.kl_multiplier_initial
elif args.kl_multiplier_schedule == 'linear':
multiplier = min(
step / (args.num_training_steps * args.kl_warmup_fraction), 1.)
return args.kl_multiplier_initial * (1. + multiplier)
# create summary writer and write to log directory
timestamp = cutils.get_timestamp()
if cutils.on_cluster():
timestamp += '||{}'.format(os.environ['SLURM_JOB_ID'])
log_dir = os.path.join(cutils.get_log_root(), args.dataset_name, timestamp)
while True:
try:
writer = SummaryWriter(log_dir=log_dir, max_queue=20)
break
except FileExistsError:
sleep(5)
filename = os.path.join(log_dir, 'config.json')
with open(filename, 'w') as file:
json.dump(vars(args), file)
best_val_elbo = -np.inf
tbar = tqdm(range(args.num_training_steps))
for step in tbar:
model.train()
optimizer.zero_grad()
scheduler.step(step)
batch = next(train_generator)[0].to(device)
elbo = model.stochastic_elbo(batch,
kl_multiplier=get_kl_multiplier(step))
loss = -torch.mean(elbo)
loss.backward()
optimizer.step()
if (step + 1) % args.monitor_interval == 0:
model.eval()
with torch.no_grad():
elbo = model.stochastic_elbo(val_batch.to(device))
mean_val_elbo = elbo.mean()
if mean_val_elbo > best_val_elbo:
best_val_elbo = mean_val_elbo
path = os.path.join(
cutils.get_checkpoint_root(),
'{}-best-val-{}.t'.format(args.dataset_name, timestamp))
torch.save(model.state_dict(), path)
writer.add_scalar(tag='val-elbo',
scalar_value=mean_val_elbo,
global_step=step)
writer.add_scalar(tag='best-val-elbo',
scalar_value=best_val_elbo,
global_step=step)
with torch.no_grad():
samples = model.sample(64)
fig, ax = plt.subplots(figsize=(10, 10))
cutils.gridimshow(make_grid(samples.view(64, 1, 28, 28), nrow=8),
ax)
writer.add_figure(tag='vae-samples', figure=fig, global_step=step)
plt.close()
# load best val model
path = os.path.join(
cutils.get_checkpoint_root(),
'{}-best-val-{}.t'.format(args.dataset_name, timestamp))
model.load_state_dict(torch.load(path))
model.eval()
np.random.seed(5)
torch.manual_seed(5)
# compute elbo on test set
with torch.no_grad():
elbo = torch.Tensor([])
log_prob_lower_bound = torch.Tensor([])
for batch in tqdm(test_loader):
elbo_ = model.stochastic_elbo(batch[0].to(device))
elbo = torch.cat([elbo, elbo_])
log_prob_lower_bound_ = model.log_prob_lower_bound(
batch[0].to(device), num_samples=1000)
log_prob_lower_bound = torch.cat(
[log_prob_lower_bound, log_prob_lower_bound_])
path = os.path.join(
log_dir, '{}-prior-{}-posterior-{}-elbo.npy'.format(
args.dataset_name, args.prior_type,
args.approximate_posterior_type))
np.save(path, utils.tensor2numpy(elbo))
path = os.path.join(
log_dir, '{}-prior-{}-posterior-{}-log-prob-lower-bound.npy'.format(
args.dataset_name, args.prior_type,
args.approximate_posterior_type))
np.save(path, utils.tensor2numpy(log_prob_lower_bound))
# save elbo and log prob lower bound
mean_elbo = elbo.mean()
std_elbo = elbo.std()
mean_log_prob_lower_bound = log_prob_lower_bound.mean()
std_log_prob_lower_bound = log_prob_lower_bound.std()
s = 'ELBO: {:.2f} +- {:.2f}, LOG PROB LOWER BOUND: {:.2f} +- {:.2f}'.format(
mean_elbo.item(), 2 * std_elbo.item() / np.sqrt(len(test_dataset)),
mean_log_prob_lower_bound.item(),
2 * std_log_prob_lower_bound.item() / np.sqrt(len(test_dataset)))
filename = os.path.join(log_dir, 'test-results.txt')
with open(filename, 'w') as file:
file.write(s)
