Maskrcnn-benchmark: ONNX model import

Created on 5 Nov 2018 · 8Comments · Source: facebookresearch/maskrcnn-benchmark

Following https://github.com/facebookresearch/maskrcnn-benchmark/issues/27, I am opening a new issue since it's not totally related.

It would be nice to be able to export the model to ONNX for later reuse with different frameworks and/or languages.

needs discussion

Source

hadim

Most helpful comment

After this PR, onnx will probably have everything we need https://github.com/onnx/onnx/pull/1010

hadim on 9 Nov 2018

👍2

All 8 comments

I think it might be difficult to have our detection models be ONNX-exportable in the near future.

The reason being that they have a few custom operators, and those operators are not in the ONNX standard.
Adding them to the ONNX standard could be possible, but I believe this requires some time and discussion between different teams. Also, having other frameworks implement those operators is yet another story.

fmassa on 6 Nov 2018

It looks like some people are already working on it:

hadim on 6 Nov 2018

👍2

That's good to know. Once ONNX supports the required ops, we can see what can be done on our side.

fmassa on 6 Nov 2018

After this PR, onnx will probably have everything we need https://github.com/onnx/onnx/pull/1010

hadim on 9 Nov 2018

👍2

hi @hadim
have u exported the model to ONNX?
can u give me some tips how to do it?
thanks

zimenglan-sysu-512 on 12 Apr 2019

https://github.com/facebookresearch/maskrcnn-benchmark/pull/138 needs to be fixed first.

hadim on 12 Apr 2019

Can the models be exported now?
The ONNX models repo contains a pre-trained Faster R-CNN model from maskrcnn-benchmark. However, it does not seem to have been exported directly.