failed to reproduce the accuracy of WRN-28-10 on Cifar-10

bmsookim / wide-resnet.pytorch

Best CIFAR-10, CIFAR-100 results with wide-residual networks using PyTorch

MIT License

462 stars 129 forks source link

failed to reproduce the accuracy of WRN-28-10 on Cifar-10 #5

Closed zhuyiche closed 5 years ago

zhuyiche commented 5 years ago

unable to reproduce the acc on cifar-10 use wrn-28-10. The gap between the reported acc and th actual acc is huge while all hyper parameters are exactly the same.

bmsookim commented 5 years ago

The problem could be with the 'F.dropout' function in the issue below!

Unfortunately I currently do not have any access to GPUs this semester :(

Will attend to the issue after June!

zhuyiche commented 5 years ago

Hi, i have checked the issues on F.dropout and altered the 'F.dropout' related codes according to https://github.com/xternalz/WideResNet-pytorch/blob/master/wideresnet.py. It terms out this is not the major problem. Seems to me there exist some implementation errors in the model. I have able to reproduce the result using xternalz git repo.