very slow training

[businiaoo]

Line 86 in the forward function in mmdet/distillation/losses/fgd.py There are two for loops here, in my test, these two for loops slow down the training seriously, is there any solution?

[dw763j]

The for loops is build upon batch size, I guess there should not be some great slowing, I mean there are just 16 samples in total.

[Muke6]

I also encountered the same problem, have you solved it?

[Tongfengyu]

Is there need to add a line ' with torch.no_grad(): ' when get the mask?