cannot reproduce the result in paper

[orangeiice]

Dear author,

Hello, I hope this message finds you well.

Thank you for sharing your work; the code is exceptionally clear and has been incredibly helpful for my own research.

I'm writing to you because I have a quick question about the results. When I ran your original code on the CIFAR dataset, I observed a performance that was about four percentage points higher than the results you reported in the paper for the corresponding experiment. I've double-checked to ensure that I haven't made any modifications to the codebase.

I was wondering if you might have any insights into what could be causing this discrepancy.

Thank you for your time and for your valuable contribution to the community.

Best regards

[tf63]

Hi @orangeiice , thank you for reaching out.

The results reported in the paper are the average over five different seeds. In my environment as well, I’ve observed that certain seeds can yield an average accuracy of around 92.5, so your result seems plausible.

That said, please note that depending on the environment, there may be slight variations—typically within ±0.1—so it’s good to keep that in mind.

Unfortunately, I no longer have access to the device I used for the experiments, so I may not be able to provide detailed support, but I hope this still helps with your investigation.

Best regards