def test_VAE(self):
model = models.VAE(
in_channels=3,
out_channels=3,
latent_channels=1,
strides=[2, 2],
hidden_channels=1,
residual_channels=1,
)
self._test_multiple_channels(model)
def test_VAE(self):
model = models.VAE(
in_channels=3,
out_channels=3,
in_size=5,
latent_dim=1,
hidden_channels=2,
n_residual_blocks=1,
residual_channels=1,
)
self._smoke_test(model, in_channels=3, test_sampling=False)
# TODO(eugenhotaj): Create a function to test VAEs sampling.
model.sample(n_images=2)
def reproduce(n_epochs=457,
batch_size=128,
log_dir="/tmp/run",
device="cuda",
debug_loader=None):
"""Training script with defaults to reproduce results.
The code inside this function is self contained and can be used as a top level
training script, e.g. by copy/pasting it into a Jupyter notebook.
Args:
n_epochs: Number of epochs to train for.
batch_size: Batch size to use for training and evaluation.
log_dir: Directory where to log trainer state and TensorBoard summaries.
device: Device to train on (either 'cuda' or 'cpu').
debug_loader: Debug DataLoader which replaces the default training and
evaluation loaders if not 'None'. Do not use unless you're writing unit
tests.
"""
from torch import optim
from torch.nn import functional as F
from torch.optim import lr_scheduler
from torch.utils import data
from torchvision import datasets
from torchvision import transforms
from pytorch_generative import trainer
from pytorch_generative import models
transform = transforms.ToTensor()
train_loader = debug_loader or data.DataLoader(
datasets.MNIST(
"/tmp/data", train=True, download=True, transform=transform),
batch_size=batch_size,
shuffle=True,
num_workers=8,
)
test_loader = debug_loader or data.DataLoader(
datasets.MNIST(
"/tmp/data", train=False, download=True, transform=transform),
batch_size=batch_size,
num_workers=8,
)
model = models.VAE(
in_channels=1,
out_channels=1,
in_size=28,
latent_dim=10,
hidden_channels=32,
n_residual_blocks=2,
residual_channels=16,
)
optimizer = optim.Adam(model.parameters(), lr=1e-3)
scheduler = lr_scheduler.MultiplicativeLR(optimizer,
lr_lambda=lambda _: 0.999977)
def loss_fn(x, _, preds):
preds, vae_loss = preds
recon_loss = F.binary_cross_entropy_with_logits(preds, x)
loss = recon_loss * 100 + vae_loss
return {
"recon_loss": recon_loss,
"vae_loss": vae_loss,
"loss": loss,
}
def sample_fn(model):
return torch.sigmoid(model.sample(n_images=64))
model_trainer = trainer.Trainer(
model=model,
loss_fn=loss_fn,
optimizer=optimizer,
train_loader=train_loader,
eval_loader=test_loader,
lr_scheduler=scheduler,
sample_epochs=5,
sample_fn=sample_fn,
log_dir=log_dir,
device=device,
)
model_trainer.interleaved_train_and_eval(n_epochs)
def reproduce(n_epochs=500,
batch_size=128,
log_dir="/tmp/run",
device="cuda",
debug_loader=None):
"""Training script with defaults to reproduce results.
The code inside this function is self contained and can be used as a top level
training script, e.g. by copy/pasting it into a Jupyter notebook.
Args:
n_epochs: Number of epochs to train for.
batch_size: Batch size to use for training and evaluation.
log_dir: Directory where to log trainer state and TensorBoard summaries.
device: Device to train on (either 'cuda' or 'cpu').
debug_loader: Debug DataLoader which replaces the default training and
evaluation loaders if not 'None'. Do not use unless you're writing unit
tests.
"""
from torch import optim
from torch.nn import functional as F
from torch.optim import lr_scheduler
from pytorch_generative import datasets
from pytorch_generative import models
from pytorch_generative import trainer
train_loader, test_loader = debug_loader, debug_loader
if train_loader is None:
train_loader, test_loader = datasets.get_mnist_loaders(
batch_size, dynamically_binarize=True, resize_to_32=True)
model = models.VAE(
in_channels=1,
out_channels=1,
latent_channels=16,
strides=[2, 2, 2, 2],
hidden_channels=64,
residual_channels=32,
)
optimizer = optim.Adam(model.parameters(), lr=5e-4)
def loss_fn(x, _, preds):
preds, kl_div = preds
recon_loss = F.binary_cross_entropy_with_logits(preds,
x,
reduction="none")
recon_loss = recon_loss.sum(dim=(1, 2, 3))
elbo = recon_loss + kl_div
return {
"recon_loss": recon_loss.mean(),
"kl_div": kl_div.mean(),
"loss": elbo.mean(),
}
def sample_fn(model):
sample = torch.sigmoid(model.sample(n_samples=16))
return torch.where(sample < 0.5, torch.zeros_like(sample),
torch.ones_like(sample))
model_trainer = trainer.Trainer(
model=model,
loss_fn=loss_fn,
optimizer=optimizer,
train_loader=train_loader,
eval_loader=test_loader,
sample_epochs=1,
sample_fn=sample_fn,
log_dir=log_dir,
device=device,
)
model_trainer.interleaved_train_and_eval(n_epochs)
