Ablation w/o multi crop?

facebookresearch / suncet

Code to reproduce the results in the FAIR research papers "Semi-Supervised Learning of Visual Features by Non-Parametrically Predicting View Assignments with Support Samples" https://arxiv.org/abs/2104.13963 and "Supervision Accelerates Pre-training in Contrastive Semi-Supervised Learning of Visual Representations" https://arxiv.org/abs/2006.10803

MIT License

486 stars 67 forks source link

Ablation w/o multi crop? #4

Closed LeeDoYup closed 3 years ago

LeeDoYup commented 3 years ago

Hello. Thanks for the awesome project ! I wonder that how the multi-crop augmentations are important for the SSL performance. Is there any results of ablation study, where the training does not contain multi-crop strategy?

MidoAssran commented 3 years ago

Hello @LeeDoYup,

We consider multi-crop an essential part of PAWS (i.e., mapping local features to global features) so we didn't experiment without using it, but if it's something you're interested in exploring I'm happy to help reviewing code/configs, or discussing results. I would expect that many of the hyper-parameters will need to be re-tuned to get the best performance possible without multi-crop.