def reproduce(n_epochs=457,
batch_size=256,
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)
model = models.PixelCNN(
in_channels=1,
out_channels=1,
n_residual=15,
residual_channels=16,
head_channels=32,
)
optimizer = optim.Adam(model.parameters(), lr=1e-3)
scheduler = lr_scheduler.MultiplicativeLR(optimizer,
lr_lambda=lambda _: 0.999977)
def loss_fn(x, _, preds):
batch_size = x.shape[0]
x, preds = x.view((batch_size, -1)), preds.view((batch_size, -1))
loss = F.binary_cross_entropy_with_logits(preds, x, reduction="none")
return loss.sum(dim=1).mean()
trainer = trainer.Trainer(
model=model,
loss_fn=loss_fn,
optimizer=optimizer,
train_loader=train_loader,
eval_loader=test_loader,
lr_scheduler=scheduler,
log_dir=log_dir,
device=device,
)
trainer.interleaved_train_and_eval(n_epochs)
def reproduce(
n_epochs=50,
batch_size=512,
log_dir="/tmp/run",
n_gpus=1,
device_id=0,
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.
n_gpus: Number of GPUs to use for training the model. If 0, uses CPU.
device_id: The device_id of the current GPU when training on multiple GPUs.
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)
model = models.NADE(input_dim=784, hidden_dim=500)
optimizer = optim.Adam(model.parameters())
def loss_fn(x, _, preds):
batch_size = x.shape[0]
x, preds = x.view((batch_size, -1)), preds.view((batch_size, -1))
loss = F.binary_cross_entropy_with_logits(preds, x, reduction="none")
return loss.sum(dim=1).mean()
model_trainer = trainer.Trainer(
model=model,
loss_fn=loss_fn,
optimizer=optimizer,
train_loader=train_loader,
eval_loader=test_loader,
log_dir=log_dir,
n_gpus=n_gpus,
device_id=device_id,
)
model_trainer.interleaved_train_and_eval(n_epochs)
def reproduce(
n_epochs=500,
batch_size=128,
log_dir="/tmp/run",
n_gpus=1,
device_id=0,
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.
n_gpus: Number of GPUs to use for training the model. If 0, uses CPU.
device_id: The device_id of the current GPU when training on multiple GPUs.
debug_loader: Debug DataLoader which replaces the default training and
evaluation loaders if not 'None'. Do not use unless you're writing unit
tests.
"""
import torch
from torch import optim
from torch.nn import functional as F
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
)
stack_configs = [
StackConfig(n_encoder_blocks=3, n_decoder_blocks=5),
StackConfig(n_encoder_blocks=3, n_decoder_blocks=5),
StackConfig(n_encoder_blocks=2, n_decoder_blocks=4),
StackConfig(n_encoder_blocks=2, n_decoder_blocks=3),
StackConfig(n_encoder_blocks=2, n_decoder_blocks=2),
StackConfig(n_encoder_blocks=1, n_decoder_blocks=1),
]
model = models.VeryDeepVAE(
in_channels=1,
out_channels=1,
input_resolution=32,
stack_configs=stack_configs,
latent_channels=16,
hidden_channels=64,
bottleneck_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,
n_gpus=n_gpus,
device_id=device_id,
)
model_trainer.interleaved_train_and_eval(n_epochs)
def reproduce(
n_epochs=50, batch_size=512, 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 import distributions
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.Compose(
[transforms.ToTensor(), lambda x: distributions.Bernoulli(probs=x).sample()]
)
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.NADE(input_dim=784, hidden_dim=500)
optimizer = optim.Adam(model.parameters())
def loss_fn(x, _, preds):
batch_size = x.shape[0]
x, preds = x.view((batch_size, -1)), preds.view((batch_size, -1))
loss = F.binary_cross_entropy_with_logits(preds, x, reduction="none")
return loss.sum(dim=1).mean()
model_trainer = trainer.Trainer(
model=model,
loss_fn=loss_fn,
optimizer=optimizer,
train_loader=train_loader,
eval_loader=test_loader,
log_dir=log_dir,
device=device,
)
model_trainer.interleaved_train_and_eval(n_epochs)
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 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_cifar10_loaders(
batch_size, normalize=True)
model = models.VQVAE2(
in_channels=3,
out_channels=3,
hidden_channels=128,
residual_channels=64,
n_residual_blocks=2,
n_embeddings=512,
embedding_dim=64,
)
optimizer = optim.Adam(model.parameters(), lr=2e-4)
scheduler = lr_scheduler.MultiplicativeLR(optimizer,
lr_lambda=lambda _: 0.999977)
def loss_fn(x, _, preds):
preds, vq_loss = preds
recon_loss = F.mse_loss(preds, x)
loss = recon_loss + 0.25 * vq_loss
return {
"vq_loss": vq_loss,
"reconstruction_loss": recon_loss,
"loss": loss,
}
model_trainer = trainer.Trainer(
model=model,
loss_fn=loss_fn,
optimizer=optimizer,
train_loader=train_loader,
eval_loader=test_loader,
lr_scheduler=scheduler,
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 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,
resize_to_32=True,
)
model = models.VeryDeepVAE(
in_channels=1,
out_channels=1,
input_resolution=32,
latent_channels=16,
hidden_channels=32,
bottleneck_channels=8,
)
optimizer = optim.Adam(model.parameters(), lr=1e-3)
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.mean(dim=(1, 2, 3))
loss = recon_loss + kl_div
return {
"recon_loss": recon_loss.mean(),
"kl_div": kl_div.mean(),
"loss": loss.mean(),
}
def sample_fn(model):
return torch.sigmoid(model.sample(n_samples=16))
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)
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.Compose([
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465),
(0.2023, 0.1994, 0.2010)),
])
train_loader = data.DataLoader(
datasets.CIFAR10("tmp/data",
train=True,
download=True,
transform=transform),
batch_size=batch_size,
shuffle=True,
pin_memory=True,
num_workers=2,
)
test_loader = data.DataLoader(
datasets.CIFAR10("tmp/data",
train=False,
download=True,
transform=transform),
batch_size=batch_size,
pin_memory=True,
num_workers=2,
)
model = models.VQVAE(
in_channels=3,
out_channels=3,
hidden_channels=128,
residual_channels=32,
n_residual_blocks=2,
n_embeddings=512,
embedding_dim=64,
)
optimizer = optim.Adam(model.parameters(), lr=2e-4)
scheduler = lr_scheduler.MultiplicativeLR(optimizer,
lr_lambda=lambda _: 0.999977)
def loss_fn(x, _, preds):
preds, vq_loss = preds
recon_loss = F.mse_loss(preds, x)
loss = recon_loss + vq_loss
return {
"vq_loss": vq_loss,
"reconstruction_loss": recon_loss,
"loss": loss,
}
model_trainer = trainer.Trainer(
model=model,
loss_fn=loss_fn,
optimizer=optimizer,
train_loader=train_loader,
eval_loader=test_loader,
lr_scheduler=scheduler,
log_dir=log_dir,
device=device,
)
model_trainer.interleaved_train_and_eval(n_epochs)
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=457,
batch_size=128,
log_dir="/tmp/run",
device="cuda",
n_channels=1,
n_pixel_snail_blocks=1,
n_residual_blocks=1,
attention_value_channels=1,
attention_key_channels=1,
evalFlag=False,
evaldir="/tmp/run",
sampling_part=1):
"""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
####################################################################################################################
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~EB~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# Load ImageGPT Data :
import gmpm
train = gmpm.train
test = gmpm.test
train_loader = data.DataLoader(
data.TensorDataset(torch.Tensor(train), torch.rand(len(train))),
batch_size=batch_size,
shuffle=True,
num_workers=8,
)
test_loader = data.DataLoader(
data.TensorDataset(torch.Tensor(test), torch.rand(len(test))),
batch_size=batch_size,
num_workers=8,
)
attention_value_channels = attention_value_channels
attention_key_channels = attention_key_channels
model = models.PixelSNAIL(
####################################################################################################################
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~EB~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# Change Input / Output size :
# 3 channels - image after clusters mapping function as input to NN :
in_channels=3,
# 512 channels - each pixel get probability to get value from 0 to 511
out_channels=512,
####################################################################################################################
n_channels=n_channels,
n_pixel_snail_blocks=n_pixel_snail_blocks,
n_residual_blocks=n_residual_blocks,
attention_value_channels=attention_value_channels,
attention_key_channels=attention_key_channels,
)
optimizer = optim.Adam(model.parameters(), lr=1e-4)
scheduler = lr_scheduler.MultiplicativeLR(optimizer,
lr_lambda=lambda _: 0.999977)
def loss_fn(x, _, preds):
####################################################################################################################
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~EB~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# Update loss function to CrossEntropyLoss :
x = x.long()
criterion = nn.NLLLoss()
B, C, D = preds.size()
preds_2d = preds.view(B, C, D, -1)
x_2d = x.view(B, D, -1)
loss = criterion(preds_2d, x_2d.long())
####################################################################################################################
return loss
_model = model.to(device)
trainer = trainer.Trainer(
model=model,
loss_fn=loss_fn,
optimizer=optimizer,
train_loader=train_loader,
eval_loader=test_loader,
lr_scheduler=scheduler,
log_dir=log_dir,
device=device,
sample_epochs=5,
sample_fn=None,
n_channels=n_channels,
n_pixel_snail_blocks=n_pixel_snail_blocks,
n_residual_blocks=n_residual_blocks,
attention_value_channels=attention_value_channels,
attention_key_channels=attention_key_channels,
evalFlag=evalFlag,
evaldir=evaldir,
sampling_part=sampling_part)
trainer.interleaved_train_and_eval(n_epochs)
