PyTorch implementation of Supervised Contrastive Learning

SupContrast: Supervised Contrastive Learning

This repo covers an reference implementation for the following papers in PyTorch, using CIFAR as an illustrative example:
(1) Supervised Contrastive Learning. Paper
(2) A Simple Framework for Contrastive Learning of Visual Representations. Paper

Loss Function

The loss function SupConLoss in losses.py takes features (L2 normalized) and labels as input, and return the loss. If labels is None or not passed to the it, it degenerates to SimCLR.

Comparison

Results on CIFAR-10:

Running

You might use CUDA_VISIBLE_DEVICES to set proper number of GPUs, and/or switch to CIFAR100 by --dataset cifar100.
(1) Standard Cross-Entropy

python main_ce.py --batch_size 1024 \
  --learning_rate 0.8 \
  --cosine --syncBN \

(2) Supervised Contrastive Learning
Pretraining stage:

python main_supcon.py --batch_size 1024 \
  --learning_rate 0.5 \
  --temp 0.5 \
  --cosine

You can also specify --syncBN but I found it not crucial for SupContrast (syncBN 95.9% v.s. BN 96.0%).
Linear evaluation stage:

python main_linear.py --batch_size 512 \
  --learning_rate 5 \
  --ckpt /path/to/model.pth

(3) SimCLR
Pretraining stage:

python main_supcon.py --batch_size 1024 \
  --learning_rate 0.5 \
  --temp 0.5 \
  --cosine --syncBN \
  --method SimCLR

The --method SimCLR flag simply stops labels from being passed to SupConLoss criterion.
Linear evaluation stage:

python main_linear.py --batch_size 512 \
  --learning_rate 1 \
  --ckpt /path/to/model.pth

t-SNE Visualization

(1) Standard Cross-Entropy

(2) Supervised Contrastive Learning

(3) SimCLR

Reference

@Article{khosla2020supervised,
    title   = {Supervised Contrastive Learning},
    author  = {Prannay Khosla and Piotr Teterwak and Chen Wang and Aaron Sarna and Yonglong Tian and Phillip Isola and Aaron Maschinot and Ce Liu and Dilip Krishnan},
    journal = {arXiv preprint arXiv:2004.11362},
    year    = {2020},
}

GitHub