Any plan to update README?

[JihwanEom]

Hello, thank you for sharing your fantastic work.

Could you please consider updating the README.md to include a comprehensive guide for training and evaluation?

Thank you!

[Zhouziyuya]

Thanks for your kind word! I have updated README and upload our pretrained weights. If you have any question don't hesitate to contact me.

[JihwanEom]

Hello, thank you for quick update! I have some questions:

1. Is it possible to share the training logs so I can verify the progress of the training locally?
2. Regarding the patch size in the code at https://github.com/jlianglab/PEAC/blob/master/utils/config.py#L15, should it be adjusted from 4 to 32? I've noticed that setting it to 32 allows the code to run smoothly. However, with the current setting of 4, there appears to be a shape mismatch error as seen at https://github.com/jlianglab/PEAC/blob/master/global_local_popar_swin.py#L173
3. Did you use single GPU when training the PEAC model with batch size 8? global_local_popar_swin.py seems like that doesn't support DDP for multi-gpu training. Could you share the required time to train swin-base w/ 448 resolution in the paper pretrained from NIHChestX-ray14?

Thanks in advance!

[Zhouziyuya]

Is it possible to share the training logs so I can verify the progress of the training locally?

I have uploaded swin_PEAC's training log please check it.

Regarding the patch size in the code at https://github.com/jlianglab/PEAC/blob/master/utils/config.py#L15, should it be adjusted from 4 to 32? I've noticed that setting it to 32 allows the code to run smoothly. However, with the current setting of 4, there appears to be a shape mismatch error as seen at https://github.com/jlianglab/PEAC/blob/master/global_local_popar_swin.py#L173

Yes, you are right! The patch size should be set as 32.

Did you use single GPU when training the PEAC model with batch size 8? global_local_popar_swin.py seems like that doesn't support DDP for multi-gpu training. Could you share the required time to train swin-base w/ 448 resolution in the paper pretrained from NIHChestX-ray14?

I tried to used DP before but I found the speed will be comparable with single GPU. The whole pretraining time is around 6 days. I used DDP for the later experiments and I will upload DDP version later.
Thanks!

[Zhouziyuya]

The ddp version code has been updated: https://github.com/jlianglab/PEAC/blob/master/global_local_popar_swin_ddp.py