PyTorch implementation of Barlow Twins.
@article{zbontar2021barlow,
title={Barlow Twins: Self-Supervised Learning via Redundancy Reduction},
author={Zbontar, Jure and Jing, Li and Misra, Ishan and LeCun, Yann and Deny, St{\'e}phane},
journal={arXiv preprint arXiv:2103.03230},
year={2021}
}
| epochs | batch size | acc1 | acc5 | download |
|---|---|---|---|---|
| 1000 | 2048 | 73.3% | 91.0% | model (logs) |
The pretrained model is also available on PyTorch Hub.
import torch
model = torch.hub.load('facebookresearch/barlowtwins:main', 'resnet50')
Install PyTorch and download ImageNet by following the instructions in the requirements section of the PyTorch ImageNet training example. The code has been developed for PyTorch version 1.7.1 and torchvision version 0.8.2, but it should work with other versions just as well.
Our best model is obtained by running the following command:
python main.py /path/to/imagenet/ --epochs 1000 --batch-size 2048 --learning-rate 0.2 --lambd 0.0051 --projector 8192-8192-8192 --scale-loss 0.024
Training time is approximately 7 days on 16 v100 GPUs.
Train a linear probe on the representations learned by Barlow Twins. Freeze the weights of the resnet and use the entire ImageNet training set.
python evaluate.py /path/to/imagenet/ /path/to/checkpoint/resnet50.pth --lr-classifier 0.1
Train a linear probe on the representations learned by Barlow Twins. Finetune the weights of the resnet and use a subset of the ImageNet training set.
python evaluate.py /path/to/imagenet/ /path/to/checkpoint/resnet50.pth --weights finetune --train-perc 1 --epochs 20 --lr-backbone 0.002 --lr-classifier 0.5 --weight-decay 0
In order to match the code that was used to develop Barlow Twins, we include an additional parameter, --scale-loss, that multiplies the loss by a constant factor. We are working on a version that will not require this parameter.
This project is under the CC-BY-NC 4.0 license. See LICENSE for details.