SimCLR

SimCLR is a self-supervised framework for visual representation learning using contrastive methods. It learns by creating two augmented views of the same image—using random cropping, color jitter, and Gaussian blur—then maximizing agreement between these augmented views while separating them from other images. Key findings include the importance of strong compositions of data augmentations, a nonlinear projection head that boosts representation quality, and the advantages of large batch sizes. Combined, these elements allow SimCLR to approach or match supervised performance on ImageNet and achieve strong transfer and semi-supervised learning results.

Key Components

  • Data Augmentations: SimCLR uses random cropping, resizing, color jittering, and Gaussian blur to create diverse views of the same image.

  • Backbone: Convolutional neural networks, such as ResNet, are employed to encode augmented images into feature representations.

  • Projection Head: A multilayer perceptron (MLP) maps features into a space where contrastive loss is applied, enhancing representation quality.

  • Contrastive Loss: The normalized temperature-scaled cross-entropy loss (NT-Xent) encourages similar pairs to align and dissimilar pairs to diverge.

Good to Know

  • Backbone Networks: SimCLR is specifically optimized for convolutional neural networks, with a focus on ResNet architectures. We do not recommend using it with transformer-based models.

  • Learning Paradigm: SimCLR is based on contrastive learning which makes it sensitive to the augmentations you pick and the method benefits from larger batch sizes.

Reference:

A Simple Framework for Contrastive Learning of Visual Representations, 2020

Tutorials:

Tutorial 3: Train SimCLR on Clothing

https://img.shields.io/badge/Open%20in%20Colab-blue?logo=googlecolab&label=%20&labelColor=5c5c5c

This example can be run from the command line with:

python lightly/examples/pytorch/simclr.py
# This example requires the following dependencies to be installed:
# pip install lightly

# 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 torch
import torchvision
from torch import nn

from lightly.loss import NTXentLoss
from lightly.models.modules import SimCLRProjectionHead
from lightly.transforms.simclr_transform import SimCLRTransform


class SimCLR(nn.Module):
    def __init__(self, backbone):
        super().__init__()
        self.backbone = backbone
        self.projection_head = SimCLRProjectionHead(512, 512, 128)

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


resnet = torchvision.models.resnet18()
backbone = nn.Sequential(*list(resnet.children())[:-1])
model = SimCLR(backbone)

device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)

transform = SimCLRTransform(input_size=32, gaussian_blur=0.0)
dataset = torchvision.datasets.CIFAR10(
    "datasets/cifar10", download=True, transform=transform
)
# or create a dataset from a folder containing images or videos:
# dataset = LightlyDataset("path/to/folder", transform=transform)

dataloader = torch.utils.data.DataLoader(
    dataset,
    batch_size=256,
    shuffle=True,
    drop_last=True,
    num_workers=8,
)

criterion = NTXentLoss()
optimizer = torch.optim.SGD(model.parameters(), lr=0.06)

print("Starting Training")
for epoch in range(10):
    total_loss = 0
    for batch in dataloader:
        x0, x1 = batch[0]
        x0 = x0.to(device)
        x1 = x1.to(device)
        z0 = model(x0)
        z1 = model(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}")