GPU and efficiency

[Sungmin-Woo]

Hi @Dwawayu, thank you for sharing the code!
Could you please provide information on the GPU(s) used for this project and the quantity?
Additionally, it would be helpful to know the training and inference runtimes.

I'm looking forward to your reply!

[Dwawayu]

Hi! Thank you for your attention! Here is some information

First stage:

Batch size: 8, 4 of which are flipped from the other 4.
Epochs: 50
Resolution: 640 * 192
GPU memory: 20G
Time: 3 days * 1 TITAN RTX, or 1 day * 4 2080Ti for me.

HR finetuning:

Batch size: 8, 4 of which are flipped from the other 4.
Epoch: 1
Resolution: 1280 * 384
GPU memory: 80G
Time: 2.3 hours * 2 A40.

Self-distillation:

Batch size: 4 without flipping
Epochs: 10
Resolution: 1280 * 384
GPU memory: 43G
Time: 14 hours * 2 A40, or 28 hours * 1 A40.

Inference:

Resolution: 1280 * 384
GPU: single 2080Ti
Time: 45 ms per frame
Time(pp): 82 ms per frame

I hope these help, please let me know if you have any other concerns.

[Sungmin-Woo]

Thank you for the detailed information! It helps a lot :)