The conversion overhead here is not negligible and it will slow down training. How about moving this to the data pipeline?

[RangiLyu]

    The conversion overhead here is not negligible and it will slow down training. How about moving this to the data pipeline?

Originally posted by @tascj in #9494 (comment)

[RangiLyu]

We are considering moving mask.to_tesnor to PackDetInputs. But this modification is related to all instance segmentation methods. We will open another PR for the refactoring.

[tascj]

Great. I tried mask.to_tesnor in PackDetInputs and it saved me 0.1x s/iter at batch size 32.

There's one related problem. You are using ndarray to store masks in BitmapMasks, which could be slow (long data_time) when used together with DataLoader. torch.Tensor with DataLoader does not have that problem.
You can speed up Mask R-CNN training a bit if using torch.Tensor in BitmapMasks. It would be nice to have that in mmdet 3.x.

[RangiLyu]

Great! Thanks for your suggestion! We will consider this feature in the refactoring.