TiCo
Example implementation of Transformation Invariance and Covariance Contrast (TiCo) for self-supervised visual representation learning. Similar to BYOL, this method is based on maximizing the agreement among embeddings of different distorted versions of the same image, which pushes the encoder to produce transformation invariant representations.
This example can be run from the command line with:
python lightly/examples/pytorch/tico.py
# Note: The model and training settings do not follow the reference settings
# from the paper. The settings are chosen such that the example can easily be
# run on a small dataset with a single GPU.
import copy
import torch
import torchvision
from torch import nn
from lightly.data import LightlyDataset
from lightly.data.multi_view_collate import MultiViewCollate
from lightly.loss.tico_loss import TiCoLoss
from lightly.models.modules.heads import TiCoProjectionHead
from lightly.models.utils import deactivate_requires_grad, update_momentum
from lightly.transforms.simclr_transform import SimCLRTransform
from lightly.utils.scheduler import cosine_schedule
class TiCo(nn.Module):
def __init__(self, backbone):
super().__init__()
self.backbone = backbone
self.projection_head = TiCoProjectionHead(512, 1024, 256)
self.backbone_momentum = copy.deepcopy(self.backbone)
self.projection_head_momentum = copy.deepcopy(self.projection_head)
deactivate_requires_grad(self.backbone_momentum)
deactivate_requires_grad(self.projection_head_momentum)
def forward(self, x):
y = self.backbone(x).flatten(start_dim=1)
z = self.projection_head(y)
return z
def forward_momentum(self, x):
y = self.backbone_momentum(x).flatten(start_dim=1)
z = self.projection_head_momentum(y)
z = z.detach()
return z
resnet = torchvision.models.resnet18()
backbone = nn.Sequential(*list(resnet.children())[:-1])
model = TiCo(backbone)
device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)
cifar10 = torchvision.datasets.CIFAR10("datasets/cifar10", download=True)
transform = SimCLRTransform(input_size=32)
dataset = LightlyDataset.from_torch_dataset(cifar10, transform=transform)
# or create a dataset from a folder containing images or videos:
# dataset = LightlyDataset("path/to/folder")
collate_fn = MultiViewCollate()
dataloader = torch.utils.data.DataLoader(
dataset,
batch_size=256,
collate_fn=collate_fn,
shuffle=True,
drop_last=True,
num_workers=8,
)
criterion = TiCoLoss()
optimizer = torch.optim.SGD(model.parameters(), lr=0.06)
epochs = 10
print("Starting Training")
for epoch in range(epochs):
total_loss = 0
momentum_val = cosine_schedule(epoch, epochs, 0.996, 1)
for (x0, x1), _, _ in dataloader:
update_momentum(model.backbone, model.backbone_momentum, m=momentum_val)
update_momentum(
model.projection_head, model.projection_head_momentum, m=momentum_val
)
x0 = x0.to(device)
x1 = x1.to(device)
z0 = model(x0)
z1 = model.forward_momentum(x1)
loss = criterion(z0, z1)
total_loss += loss.detach()
loss.backward()
optimizer.step()
optimizer.zero_grad()
avg_loss = total_loss / len(dataloader)
print(f"epoch: {epoch:>02}, loss: {avg_loss:.5f}")
This example can be run from the command line with:
python lightly/examples/pytorch_lightning/tico.py
import copy
import pytorch_lightning as pl
import torch
import torchvision
from torch import nn
from lightly.data import LightlyDataset
from lightly.data.multi_view_collate import MultiViewCollate
from lightly.loss.tico_loss import TiCoLoss
from lightly.models.modules.heads import TiCoProjectionHead
from lightly.models.utils import deactivate_requires_grad, update_momentum
from lightly.transforms.simclr_transform import SimCLRTransform
from lightly.utils.scheduler import cosine_schedule
class TiCo(pl.LightningModule):
def __init__(self):
super().__init__()
resnet = torchvision.models.resnet18()
self.backbone = nn.Sequential(*list(resnet.children())[:-1])
self.projection_head = TiCoProjectionHead(512, 1024, 256)
self.backbone_momentum = copy.deepcopy(self.backbone)
self.projection_head_momentum = copy.deepcopy(self.projection_head)
deactivate_requires_grad(self.backbone_momentum)
deactivate_requires_grad(self.projection_head_momentum)
self.criterion = TiCoLoss()
def forward(self, x):
y = self.backbone(x).flatten(start_dim=1)
z = self.projection_head(y)
return z
def forward_momentum(self, x):
y = self.backbone_momentum(x).flatten(start_dim=1)
z = self.projection_head_momentum(y)
z = z.detach()
return z
def training_step(self, batch, batch_idx):
(x0, x1), _, _ = batch
momentum = cosine_schedule(self.current_epoch, 10, 0.996, 1)
update_momentum(self.backbone, self.backbone_momentum, m=momentum)
update_momentum(self.projection_head, self.projection_head_momentum, m=momentum)
x0 = x0.to(self.device)
x1 = x1.to(self.device)
z0 = self.forward(x0)
z1 = self.forward_momentum(x1)
loss = self.criterion(z0, z1)
return loss
def configure_optimizers(self):
return torch.optim.SGD(self.parameters(), lr=0.06)
model = TiCo()
cifar10 = torchvision.datasets.CIFAR10("datasets/cifar10", download=True)
transform = SimCLRTransform(input_size=32)
dataset = LightlyDataset.from_torch_dataset(cifar10, transform=transform)
# or create a dataset from a folder containing images or videos:
# dataset = LightlyDataset("path/to/folder")
collate_fn = MultiViewCollate()
dataloader = torch.utils.data.DataLoader(
dataset,
batch_size=256,
collate_fn=collate_fn,
shuffle=True,
drop_last=True,
num_workers=8,
)
accelerator = "gpu" if torch.cuda.is_available() else "cpu"
trainer = pl.Trainer(max_epochs=10, devices=1, accelerator=accelerator)
trainer.fit(model=model, train_dataloaders=dataloader)
This example runs on multiple gpus using Distributed Data Parallel (DDP) training with Pytorch Lightning. At least one GPU must be available on the system. The example can be run from the command line with:
python lightly/examples/pytorch_lightning_distributed/tico.py
The model differs in the following ways from the non-distributed implementation:
Distributed Data Parallel is enabled
Synchronized Batch Norm is used in place of standard Batch Norm
Note that Synchronized Batch Norm is optional and the model can also be trained without it. Without Synchronized Batch Norm the batch norm for each GPU is only calculated based on the features on that specific GPU.
import copy
import pytorch_lightning as pl
import torch
import torchvision
from torch import nn
from lightly.data import LightlyDataset
from lightly.data.multi_view_collate import MultiViewCollate
from lightly.loss.tico_loss import TiCoLoss
from lightly.models.modules.heads import TiCoProjectionHead
from lightly.models.utils import deactivate_requires_grad, update_momentum
from lightly.transforms.simclr_transform import SimCLRTransform
from lightly.utils.scheduler import cosine_schedule
class TiCo(pl.LightningModule):
def __init__(self):
super().__init__()
resnet = torchvision.models.resnet18()
self.backbone = nn.Sequential(*list(resnet.children())[:-1])
self.projection_head = TiCoProjectionHead(512, 1024, 256)
self.backbone_momentum = copy.deepcopy(self.backbone)
self.projection_head_momentum = copy.deepcopy(self.projection_head)
deactivate_requires_grad(self.backbone_momentum)
deactivate_requires_grad(self.projection_head_momentum)
self.criterion = TiCoLoss()
def forward(self, x):
y = self.backbone(x).flatten(start_dim=1)
z = self.projection_head(y)
return z
def forward_momentum(self, x):
y = self.backbone_momentum(x).flatten(start_dim=1)
z = self.projection_head_momentum(y)
z = z.detach()
return z
def training_step(self, batch, batch_idx):
(x0, x1), _, _ = batch
momentum = cosine_schedule(self.current_epoch, 10, 0.996, 1)
update_momentum(self.backbone, self.backbone_momentum, m=momentum)
update_momentum(self.projection_head, self.projection_head_momentum, m=momentum)
x0 = x0.to(self.device)
x1 = x1.to(self.device)
z0 = self.forward(x0)
z1 = self.forward_momentum(x1)
loss = self.criterion(z0, z1)
return loss
def configure_optimizers(self):
return torch.optim.SGD(self.parameters(), lr=0.06)
model = TiCo()
cifar10 = torchvision.datasets.CIFAR10("datasets/cifar10", download=True)
transform = SimCLRTransform(input_size=32)
dataset = LightlyDataset.from_torch_dataset(cifar10, transform=transform)
# or create a dataset from a folder containing images or videos:
# dataset = LightlyDataset("path/to/folder")
collate_fn = MultiViewCollate()
dataloader = torch.utils.data.DataLoader(
dataset,
batch_size=256,
collate_fn=collate_fn,
shuffle=True,
drop_last=True,
num_workers=8,
)
# Train with DDP and use Synchronized Batch Norm for a more accurate batch norm
# calculation. Distributed sampling is also enabled with replace_sampler_ddp=True.
trainer = pl.Trainer(
max_epochs=10,
devices="auto",
accelerator="gpu",
strategy="ddp",
sync_batchnorm=True,
use_distributed_sampler=True, # or replace_sampler_ddp=True for PyTorch Lightning <2.0
)
trainer.fit(model=model, train_dataloaders=dataloader)