ReScience
[Re] Bootstrap Your Own Latent: A new approach to self-supervised learning
Original article: J.-B. Grill, F. Strub, F. Altché, C. Tallec, P. H. Richemond, E. Buchatskaya, C. Doersch, B. A. Pires, Z. D. Guo, and M. G. Azar. “Bootstrap your own latent: A new approach to self-supervised learning.” In: arXiv preprint arXiv:2006.07733 (2020)
PDF URL: https://github.com/ADevillers/BYOL/blob/main/report.pdf Metadata URL: https://github.com/ADevillers/BYOL/blob/main/report.metadata.tex Code URL: https://github.com/ADevillers/BYOL/tree/main
Scientific domain: Representation Learning Programming language: Python Suggested editor: @rougier
Thansk for your submission and sorry for the delay, we'll assign an editor soon.
I can handle this one too (I'll try not to mix up between the two consecutive submissions...).
@ADevillers : given the fact that you results were obtained using HPC, that the reviewers may not have access to, are there ways to test you code without it (like running only a subpart of it...)?
(same question applies to the companion paper)
Yes, you can train on CIFAR10 or perform linear evaluation on CIFAR10 and ImageNet with the following setup:
- Ensure you use
--computer='other'. - For GPU usage (recommended for faster processing), set
--hardware='mono-gpu'. If you're using a CPU, set--hardware='cpu'. - Follow the commands provided in the README's final section.
Note: Training on ImageNet without HPC is impractical due to the large memory requirements for the necessary batch size and the extended duration it might take; yet, evaluation may be fine.
Good news: @charlypg has accepted to review this paper and its companion!