ResNet-20 accuracy on CIFAR-10

AngusG / bn-advex-zhang-fixup

Code for ResNet-Fixup experiments as part of "Batch Norm is a Cause of Adversarial Vulnerability" presented at http://deep-phenomena.org/

https://arxiv.org/abs/1905.02161

BSD 3-Clause "New" or "Revised" License

3 stars 0 forks source link

ResNet-20 accuracy on CIFAR-10 #1

Open skvenka5 opened 5 years ago

skvenka5 commented 5 years ago

Hey, I am trying to reproduce resnet20 fixup results given in "fixup-vs-batchnorm-lab-clean-eps-converted-to.pdf". But, after 200 epochs I am getting only ~35% accuracy on CIFAR-10. I ran the model with default hyperparameters using the following command CUDA_VISIBLE_DEVICES=0 python cifar_train.py -a fixup_resnet20 What could be the issue?

AngusG commented 4 years ago

Hi, sorry I didn't notice this issue until now.

Are you still stuck with this? If so can you paste the output of the first few training epochs here? And maybe also your pytorch version?

I just ran python cifar_train.py --dataroot /path/to/my/dataset/ --logdir /log/path/ -a fixup_resnet20 --sess skvenka5 and have ~80% test accuracy after only 15 epochs.

The best accuracy is 91.66% @ epoch 191, compatible with the "Fixup" data series of figure you mention.