HobbitLong / SupContrast

PyTorch implementation of "Supervised Contrastive Learning" (and SimCLR incidentally)
BSD 2-Clause "Simplified" License
3.1k stars 532 forks source link

About train resent-18 on the dataset CIFAR-100, the accuracy is just 55. #113

Open TouchNow opened 2 years ago

TouchNow commented 2 years ago

I have ran 'main_supcon.py'. Parameter: CIFAR-100 resnet-18, other parameters have no change. Then I ran 'main_linear' to test resnet-18's accuracy(Load the encoder weights previously trained in 'main_supcon.py' ), just train the last layer(fc layer). But the accuracy is about 55(Baseline is resent-18 train alone, the accuracy is 70). I think it is strange, so want to consult you about it.

shuaiNJU commented 1 year ago

Hi, have you found the reason?