Very long training time

Open david-az opened this issue 4 years ago • 1 comments

Hi, thank you for your great work.

The number of FLOPs and the numbers of parameters are less than Swin Transformer, however the training time of HRFormer is at least 2 times longer than Swin, and 3 times longer than HRNet. I guess the gradient calculation is very long because of a lot of reshape operations ? Is there a way to optimize that ?

Thank you

Dec 07 '21 10:12 david-az

@david-az Good question!

Currently, we do not have any plans or solutions to optimize the training time cost of HRFormer. You can find that our HRFormer already chooses a much smaller network depth compared to the original HRNet. Any suggestions on optimizing the training time are WELCOME!

Dec 08 '21 14:12 PkuRainBow