Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Implement torchvision.ops.roi_align in torchxla2 #8288

Open
qihqi opened this issue Oct 18, 2024 · 0 comments · May be fixed by #8362
Open

Implement torchvision.ops.roi_align in torchxla2 #8288

qihqi opened this issue Oct 18, 2024 · 0 comments · May be fixed by #8362
Assignees

Comments

@qihqi
Copy link
Collaborator

qihqi commented Oct 18, 2024

🚀 Feature

https://pytorch.org/vision/stable/generated/torchvision.ops.roi_align.html?highlight=roi_align#torchvision.ops.roi_align

Few ideas:

  1. Use torch decomposition in here: https://github.com/pytorch/vision/blob/main/torchvision/ops/roi_align.py#L115 ; tried this and found out jax OOMs pointing here: https://github.com/pytorch/vision/blob/main/torchvision/ops/roi_align.py#L74 so the issue seems that the advanced indexing used here creates large intermediaries. Torch side needed a "loop-less" impl to help with inductor, we could actually rewrite it using jax.vmap and jax.lax.fori_loop.
  2. Start from this jax implementation: https://github.com/google-research/scenic/blob/74225e8e71ba27a76abd62e6bc56e8a64c4cc19e/scenic/projects/baselines/centernet/modeling/roi_align.py#L103 but this one takes output_size as int instead of tuple of int (i.e. it assumes width and height is the same) so it will need some modification.

Motivation

Pitch

Alternatives

Additional context

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging a pull request may close this issue.

1 participant