Reproducibility Gap in CIFAR-10 Experiments

[zyrakael]

Thank you very much for sharing your work and providing the code. I truly appreciate your contributions to this field.

After strictly following the provided code and instructions, the performance on CIFAR-10 is significantly worse than the results reported in the original paper. Could you confirm whether the shared code is identical to the implementation used in your experiments? If not, are there critical implementation details or hyperparameter adjustments that might explain this discrepancy?

I would be sincerely grateful for any insights you could provide. Thank you again for your time and for making your research available!

[OngWinKent]

Hi Wang,

Thank you for your interest in this work.

Controlling hyperparameter sensitivity, particularly for batch_size and learning_rate, is our next step. We are researching strategies to enhance stability and ensure consistent performance across different models with minimal tuning.

If you have any suggestions, we welcome your insights and would love to explore potential collaborations.