SimCLR
A Simple Framework for Contrastive Learning of Visual Representations
Ting Chen, Simon Kornblith, Mohammad Norouzi, Geoffrey Hinton · ICML 2020
Implement NT-Xent over 2B anchors with positive-pair mapping, self masking, same-view negatives and temperature.
What you will implement
Normalized contrastive loss, independent gradient checks and actual CPU training on 6 vector pairs; no image-training or paper benchmark reproduction.
- 84. SimCLR Normalized Temperature Contrastive Loss
Implement NT-Xent over 2B anchors including same-view negatives, preserving gradients into both input views.
Implement only NT-Xent; a supplied CPU lab trains 6 vector pairs, without an image encoder or paper benchmark reproduction.