Some Confusion About Model Config Files

[Pockern]

Thanks for your excellent work!

I'm reproducing your experiments and following the code, but I have some questions regarding the configuration in train_ERR_UHDLL.yml:

1. Learning Rate Schedule: The paper mentions training with AdamW, starting at LR 0.0005 and reducing to 1e-7 via cosine annealing over 100k iterations. However, the config file shows the LR being reduced from 5e-4 to 1e-4 over 500k iterations.
2. Training Patch Size: The paper specifies a patch size of 512x512 for UHDLL, while the config uses 768x768.

Could you please advise on whether I should use the settings in the config file or the parameters described in the paper? Any guidance would be greatly appreciated!

[ChenzhaoNju]

可以以我代码放出来的为准，最后比较实验采用相同的setting就可以