Reproducibility Issues

[p-lam]

Hello,

I am unable to produce the results on your paper for CIFAR-100 IF 100 with a Resnet-32 backbone. I have followed the details in the paper (cosine scheduler, lr=0.01, batch size 64, etc) and made some minor modifications to your code to take in a Resnet-32 (see resnet_cifar.py in MiSLAS). However, even with weight decay tuning via bayesian optimization, I cannot attain a test accuracy beyond 39.5% (small improvement from baseline).

Would you be able to share the specific weight decay value used to attain the paper results (46.08% acc)?

Thank you for your time!