VICReg: Variance-Invariance-Covariance Regularization For Self-Supervised Learning

This repository provides a PyTorch implementation and pretrained models for VICReg, as described in the paper VICReg: Variance-Invariance-Covariance Regularization For Self-Supervised Learning
Adrien Bardes, Jean Ponce and Yann LeCun
Meta AI, Inria

Pre-trained Models

You can choose to download only the weights of the pretrained backbone used for downstream tasks, or the full checkpoint which contains backbone and projection head weights.

arch	params	accuracy	download
ResNet-50	23M	73.2%	backbone only	full ckpt
ResNet-50 (x2)	93M	75.5%	backbone only	full ckpt
ResNet-200 (x2)	250M	77.3%	backbone only	full ckpt

Pretrained models on PyTorch Hub

import torch
resnet50 = torch.hub.load('facebookresearch/vicreg:main', 'resnet50')
resnet50x2 = torch.hub.load('facebookresearch/vicreg:main', 'resnet50x2')
resnet200x2 = torch.hub.load('facebookresearch/vicreg:main', 'resnet200x2')

Training

Install PyTorch (pytorch.org) and download ImageNet. The code has been developed for PyTorch version 1.8.1 and torchvision version 0.9.1, but should work with other versions just as well.

Single-node local training

To pretrain VICReg with MobileNetV2 on a single node with 4 GPUs for 100 epochs, run:

python -m torch.distributed.launch --nproc_per_node=4 main_vicreg.py --data-dir /scratch/datasets/imagenet/ --exp-dir /path/to/experiment/ --arch mobilenetv2 --epochs 100 --batch-size 256 --base-lr 0.3

Multi-node training with SLURM (DGX)

To pretrain VICReg with submitit (pip install submitit) and SLURM on 4 nodes with 8 GPUs each for 1000 epochs, run:

python run_with_submitit.py --nodes 2 --ngpus 8 --data-dir $DATA_ILSVRC2012 --exp-dir experiment_slurm --arch mobilenetv2 --epochs 1000 --batch-size 2048 --base-lr 0.2

Evaluation

Linear evaluation

To evaluate a pretrained MobileNetV2 backbone on linear classification on ImageNet, run:

python evaluate.py --data-dir /scratch/datasets/imagenet/ --pretrained /path/to/checkpoint/mobilenetv2.pth --exp-dir /path/to/experiment/ --lr-head 0.02

Semi-supervised evaluation

To evaluate a pretrained ResNet50-model on semi-supervised fine-tunning on 1% of ImageNet labels, run:

python evaluate.py --data-dir /path/to/imagenet/ --pretrained /path/to/checkpoint/resnet50.pth --exp-dir /path/to/experiment/ --weights finetune --train-perc 1 --epochs 20 --lr-backbone 0.03 --lr-classifier 0.08 --weight-decay 0

To evaluate a pretrained ResNet50-model on semi-supervised fine-tunning on 10% of ImageNet labels, run:

python evaluate.py --data-dir /path/to/imagenet/ --pretrained /path/to/checkpoint/resnet50.pth --exp-dir /path/to/checkpoint/resnet.pth --weights finetune --train-perc 10 --epochs 20 --lr-backbone 0.01 --lr-classifier 0.1 --weight-decay 0

Acknowledgement

This repository is built using the Barlow Twins repository.

License

This project is released under MIT License, which allows commercial use. See LICENSE for details.

Citation

If you find this repository useful, please consider giving a star ⭐ and citation:

@inproceedings{bardes2022vicreg,
  author  = {Adrien Bardes and Jean Ponce and Yann LeCun},
  title   = {VICReg: Variance-Invariance-Covariance Regularization For Self-Supervised Learning},
  booktitle = {ICLR},
  year    = {2022},
}

FrancescoConti/vicreg