Reproducibility Issues
Opened this issue · 2 comments
entuo commented
Hello, this is great work. Congratulations! But I have a question: when I run the code, I find that the experimental results for CIFAR-10-LT and CIFAR-100-LT do not reach the accuracy mentioned in your paper. What could be the possible reasons for this, and are there any details I might have overlooked? Thank you for your response.
Keke921 commented
There may be some hyperparameters that need to be adjusted. For example, batch size, learning rate, and m. And I recently found that the PyTorch version and hardware environment also have a significant impact on the final performance.
在 2024年6月1日,17:39,entuo ***@***.***> 写道:
Hello, this is great work. Congratulations! But I have a question: when I run the code, I find that the experimental results for CIFAR-10-LT and CIFAR-100-LT do not reach the accuracy mentioned in your paper. What could be the possible reasons for this, and are there any details I might have overlooked? Thank you for your response.
—
Reply to this email directly, view it on GitHub<#1>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/ALDW7LFRI72X44ZE4JBDUL3ZFGJDHAVCNFSM6AAAAABIUAKUK6VHI2DSMVQWIX3LMV43ASLTON2WKOZSGMZDSMBTHE4TQOA>.
You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>
entuo commented
Thank you for your answer. What you mean is that different devices have different intrinsically random seeds even though they have the same random seed number, which will have a big impact on the results.