[feature request] ROI Pooling layers

[varunagrawal]

It would be great to have support for various ROI Pooling operations as easy to add layers to facilitate research in object detection and semantic/instance segmentation.

Here is a live checklist:

• ROI Pooling Support for ROI Pooling #592 ROIPool: Support for all datatypes #632
• Position Specific ROI Pooling
• ROI Align Support for ROIAlign Layer #630

General PRs: #626

[fmassa]

I agree. I've started sketching the structure of it in https://github.com/pytorch/vision/tree/layers?files=1 .
I'll look into opening a PR tomorrow with a few layers

[wadimkehl]

Any movement on this?

[fmassa]

Hey Wadim,
ROIPool and ROIAlign are implemented in the layers branch. I'm holding on merging them as is because I might want to change a few things, but feel free to use them as is (they are working)

[wadimkehl]

Great, will have a look! Thanks :)

[varunagrawal]

@fmassa any updates on this? I'm sure a lot of people would benefit from having a master branch version of this available soon.

[botcs]

It would be super convenient to have this installed automatically with torch/torchvision

[fmassa]

Having the master branch have cpu/cuda layers officially requires a few additional changes, like providing wheels with the compiled binaries for each supported architecture, and I'm not looking at this at the moment.

[rawmarshmellows]

Just wondering if the ROI pooling/align could theoretically be done in pure Pytorch (even if it will be slow?)

[botcs]

Was thinking about the same...
ROI pooling: Adaptive MaxPools exist, if you can efficiently crop out all the tensors you need from each image in a batch and concat them in the batch dimension, maybe it could work, however I have a bad feeling about the efficiency of this naive approach

[fmassa]

@kevinlu1211 it is possible to implement it using pure PyTorch, and performance is OK.
An (old, badly tested) implementation can be found in https://github.com/pytorch/examples/pull/21/files#diff-7573d025c4128229f8efa3ff042e09d1R38

[rawmarshmellows]

You are a life saver! I’m just writing a tutorial to explain mask rcnn thanks a lot!

…

On Tue, 19 Jun 2018 at 11:44 pm, Francisco Massa ***@***.***> wrote: @kevinlu1211 <https://github.com/kevinlu1211> it is possible to implement it using pure PyTorch, and performance is OK. An (old, badly tested) implementation can be found in https://github.com/pytorch/examples/pull/21/files#diff-7573d025c4128229f8efa3ff042e09d1R38 — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#477 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AME8D-m7bWhfCgW9yzr5VXr8fwUFcI4Sks5t-QAkgaJpZM4Tifuk> .

[varunagrawal]

@fmassa I find it surprising this is not higher priority. The fact that every other major deep learning framework supports ROI Pooling and there is no easy way to write a Pytorch version of Detectron for research purposes despite the deep integration between Pytorch and Caffe2 is bewildering.

Is there some other way we can push this forward if you're too busy? I'm sure we can find volunteers to push this out the door as soon as possible.

[wadimkehl]

Come on @fmassa, make us all happy. If you don't have time, I'd gladly help!

[vfdev-5]

Yeah, captain @fmassa you have almost an army of volunteers that wait for your orders :)