Self-supervised learning

Lightly SSL is a computer vision framework for training deep learning models using self-supervised learning. The framework can be used for a wide range of useful applications such as finding the nearest neighbors, similarity search, transfer learning, or data analytics.

How Lightly SSL Works

The flexible design of Lightly SSL makes it easy to integrate in your Python code. Lightly SSL is built completely around PyTorch and the different pieces can be put together to fit your requirements.

Data and Transformations

The basic building block of contrastive self-supervised methods such as SimCLR are image transformations. Each image is transformed into several new images by randomly applied augmentations. The task of the self-supervised model is then to identify the images which come from the same original among a set of negative examples.

For example, the transforms below will apply the SimCLR image transform to the input images.

from lightly.transforms.simclr_transform import SimCLRTransform

# The following transform will return two augmented images per input image.
transform = SimCLRTransform()

Let’s now load an image dataset and create a PyTorch dataloader.

import torch

# Create a dataset from your image folder.
dataset = data.LightlyDataset(
    input_dir='./my/cute/cats/dataset/',
    transform=transform,
)

# Build a PyTorch dataloader.
dataloader = torch.utils.data.DataLoader(
    dataset,                # Pass the dataset to the dataloader.
    batch_size=128,         # A large batch size helps with learning.
    shuffle=True,           # Shuffling is important!
)

Note

You can also use a custom PyTorch Dataset instead of the LightlyDataset. Just make sure your Dataset implementation returns a tuple of (sample, target, filename) to support the basic functions for training models. See lightly.data.dataset for more information.

Head to the next section to see how you can train a ResNet on the data you just prepared.

Model, Loss and Training

Now, we need an embedding model, an optimizer and a loss function. We use a ResNet together with the normalized temperature-scaled cross entropy loss and simple stochastic gradient descent.

import torchvision

from lightly.loss import NTXentLoss
from lightly.models.modules.heads import SimCLRProjectionHead

# use a resnet backbone
resnet = torchvision.models.resnet18()
resnet = torch.nn.Sequential(*list(resnet.children())[:-1])

# build a SimCLR model
class SimCLR(torch.nn.Module):
    def __init__(self, backbone, hidden_dim, out_dim):
        super().__init__()
        self.backbone = backbone
        self.projection_head = SimCLRProjectionHead(hidden_dim, hidden_dim, out_dim)

    def forward(self, x):
        h = self.backbone(x).flatten(start_dim=1)
        z = self.projection_head(h)
        return z

model = SimCLR(resnet, hidden_dim=512, out_dim=128)

# use a criterion for self-supervised learning
# (normalized temperature-scaled cross entropy loss)
criterion = NTXentLoss(temperature=0.5)

# get a PyTorch optimizer
optimizer = torch.optim.SGD(model.parameters(), lr=1e-0, weight_decay=1e-5)

Note

You can also use custom backbones and use lightly to train them using self-supervised learning. Learn more about how to use custom backbones in our colab playground.

Train the model for 10 epochs.

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
max_epochs = 10
for epoch in range(max_epochs):
    for (x0, x1), _, _ in dataloader:

        x0 = x0.to(device)
        x1 = x1.to(device)

        z0 = model(x0)
        z1 = model(x1)

        loss = criterion(z0, z1)
        loss.backward()

        optimizer.step()
        optimizer.zero_grad()

Congrats, you just trained your first model using self-supervised learning!

You can of course also use PyTorch Lightning to implement and train your model.

import pytorch_lightning as pl

class SimCLR(pl.LightningModule):
    def __init__(self, backbone, hidden_dim, out_dim):
        super().__init__()
        self.backbone = backbone
        self.projection_head = SimCLRProjectionHead(hidden_dim, hidden_dim, out_dim)
        self.criterion = NTXentLoss(temperature=0.5)

    def forward(self, x):
        h = self.backbone(x).flatten(start_dim=1)
        z = self.projection_head(h)
        return z

    def training_step(self, batch, batch_idx):
        (x0, x1), _, _ = batch
        z0 = self.forward(x0)
        z1 = self.forward(x1)
        loss = self.criterion(z0, z1)
        return loss

    def configure_optimizers(self):
        optimizer = torch.optim.SGD(self.parameters(), lr=1e-0)
        return optimizer

model = SimCLR(resnet, hidden_dim=512, out_dim=128)
trainer = pl.Trainer(max_epochs=max_epochs, devices=1, accelerator="gpu")
trainer.fit(
    model,
    dataloader
)

To train on a machine with multiple GPUs we recommend using the distributed data parallel strategy.

# If we have a machine with 4 GPUs we set devices=4 and accelerator="gpu".
trainer = pl.Trainer(
    max_epochs=max_epochs,
    devices=4,
    accelerator="gpu",
    strategy='ddp'
)
trainer.fit(
    model,
    dataloader
)

Embeddings

You can use the trained model to embed your images or even access the embedding model directly.

# make a new dataloader without the transformations
# The only transformation needed is to make a torch tensor out of the PIL image
dataset.transform = torchvision.transforms.ToTensor()
dataloader = torch.utils.data.DataLoader(
    dataset,        # use the same dataset as before
    batch_size=1,   # we can use batch size 1 for inference
    shuffle=False,  # don't shuffle your data during inference
)

# embed your image dataset
embeddings = []
model.eval()
with torch.no_grad():
    for img, label, fnames in dataloader:
        img = img.to(model.device)
        emb = model.backbone(img).flatten(start_dim=1)
        embeddings.append(emb)

    embeddings = torch.cat(embeddings, 0)

Done! You can continue to use the embeddings to find nearest neighbors or do similarity search. Furthermore, the ResNet backbone can be used for transfer and few-shot learning.

# access the ResNet backbone
resnet = model.backbone

Note

Self-supervised learning does not require labels for a model to be trained on. Lightly, however, supports the use of additional labels. For example, if you train a model on a folder ‘cats’ with subfolders ‘Maine Coon’, ‘Bengal’ and ‘British Shorthair’ Lightly SSL automatically returns the enumerated labels as a list.

Lightly SSL in Three Lines

Lightly SSL also offers an easy-to-use interface. The following lines show how the package can be used to train a model with self-supervision and create embeddings with only three lines of code.

from lightly.core import train_embedding_model, embed_images

# first we train our model for 10 epochs
checkpoint = train_embedding_model(input_dir='./my/cute/cats/dataset/', trainer={'max_epochs': 10})

# let's embed our 'cats' dataset using our trained model
embeddings, labels, filenames = embed_images(input_dir='./my/cute/cats/dataset/', checkpoint=checkpoint)

# now, let's inspect the shape of our embeddings
print(embeddings.shape)