barlowtwins

PyTorch Implementation of Barlow Twins paper: Barlow Twins: Self-Supervised Learning via Redundancy Reduction

This is currently a work in progress. The code is a modified version of the SimSiam implementation here

• Time per epoch is around 40 seconds on a V100 GPU
• GPU usage is around 9 GBytes
• The current version reaches around 84.7% test accuracy

Todo:

• warmup learning rate from 0
• report results on cifar-10
• create PR to add to lightly

Installation

pip install -r requirements.txt

Dependencies

• PyTorch
• PyTorch Lightning
• Torchvision
• lightly

Benchmarks

We benchmark the BarlowTwins model on the CIFAR-10 dataset following the KNN evaluation protocol. Currently, the best effort achieved a test accuracy of 84.7%.

Accuracy	Loss

Paper

Barlow Twins: Self-Supervised Learning via Redundancy Reduction