/SupCon

PyTorch implementation of SupCon (Supervised Contrastive Learning) and SimCRL (A Simple Framework for Contrastive Learning of Visual Representations)

Primary LanguagePython

Because Mnist is too simple, we exchange training sets and test sets to make it more difficult. All experiments are performed 5 times and we report the average of the accuracy.

official implementation:

Loss Test Accuracy
CrossEntropy 0.9783
SimCLR 0.9460
SupCon 0.9772

our implementation:

Loss Test Accuracy
CrossEntropy 0.9783
SimCLR 0.9475
SupCon 0.9769