search

lyakaap / VAT-pytorch

Virtual Adversarial Training (VAT) implementation for PyTorch
296 stars 43 forks source link

Test Accuracy #3

Open mfaraki opened 6 years ago

mfaraki commented 6 years ago

Hi, Thanks for sharing. Are the hyper-parameters set optimally? I get around 50% test accuracy on the CIFAR10 dataset, which I assume is very low.

Thanks

sndnyang commented 5 years ago

I think the hyper-parameters are not optimally. I got 58% test accuracy on cifar10. Using another code with the same arch CNN(the hyper-parameters are really different), I got around 70%.

sndnyang commented 5 years ago

By the way, the arch used is much simpler than the one in the paper.

luzai commented 5 years ago

May I ask the optimal hyperparameter? I try with xi=1e-6, eps=8, batch_size=128, performance is better, the test acc is 60.74%.

LesiChou commented 4 years ago

Hi! But I found that If you remove the VAT, you could get a similar results.