search

pytorch / vision

Datasets, Transforms and Models specific to Computer Vision
https://pytorch.org/vision
BSD 3-Clause "New" or "Revised" License
16.18k stars 6.95k forks source link

DeformConv2d offset and mask backprop #5989

Open Cypher30 opened 2 years ago

Cypher30 commented 2 years ago

🚀 The feature

I notice that the current vision of DeformConv2d module doesn’t treat offset and mask as parameters, so it seems that the standard backward routine cannot update the offset. Is there any possibility that you could add this feature in the future?

Motivation, pitch

I’m trying to build DNNs with above module, so it is not that convenient if I need to rewrite the backward process by myself

Alternatives

No response

Additional context

No response

datumbox commented 2 years ago

Thanks for the feature request. We can look into doing this work on the future. It's not something we can pick up immediately because of lack of bandwidth, so note that you might need to find your own workarounds until we pick this up.

Cypher30 commented 2 years ago

Thanks a lot for your reply!

Zhaoyi-Yan commented 2 years ago

@Cypher30 Maybe this work https://arxiv.org/abs/2107.03665 can help.

Cypher30 commented 2 years ago

@Cypher30 Maybe this work https://arxiv.org/abs/2107.03665 can help.

Thanks a lot!

RGring commented 1 year ago

+1