Training Details for PatchCraft？

[yjtlab]

Thanks for the great work! I reproduced results with your pretrained weights, but my training fails to match the paper's metrics. Could you clarify:

​Augmentation: Exact RandAugment/JPEG compression settings?
​Optimization: EMA/LR warmup used? How to balance loss terms?
​Critical Tricks: Gradient clipping, layer freezing, or mixed precision?
Detailed training configs would help the community. Thanks