Is there any demo on how to use d2go mask-rcnn model on android?

It seems like there are still some issues regarding deploying mask-rcnn on android, does anyone managed to solve this ?

Hey, I think the d2go demo is already available. Are you able to find it? Or you have some difficulties to setup the demo? Can you help clarify the the question so that we can better support you.
@jeffxtang may be helpful to give your more detailed instruction once you are able to clarify the issue you are facing.

FYI. In case you can not find the demo, D2go demo for android with instruction at android-demo-app/D2Go at master · pytorch/android-demo-app · GitHub.

the demo is deploying a faster-rcnn model, but when trying to do the same steps to deploy a mask-rcnn model, I got some errors like “AssertionError: Per channel weight observer is not supported yet for ConvTranspose{n}d.” error. This error comes when trying to export the model to create a .pt file to be used on mobile.
Fixing this issue by walking around and using using d2go export scripts yields another error when loading the model on android

Unknown builtin op: _caffe2::GenerateProposals. Could not find any similar ops to _caffe2::GenerateProposals. This op may not exist or may not be currently supported in TorchScript.

What do I miss ? is there a demo to deploy a mask-rcnn segmentation on android ?

Thanks for the update! Same issue is reported at Using D2Go's mask_rcnn model in the demo app · Issue #121 · pytorch/android-demo-app · GitHub. We will let you know once we get a solution for it. CC: @jeffxtang @fmassa

Please see the update by @zhanghang1989 at Using D2Go's mask_rcnn model in the demo app · Issue #121 · pytorch/android-demo-app · GitHub. Also - “exporting mask rcnn is a known issue. We are updating the code and need to retrain the models. It should be fixed later this week.”

1 Like

Any updates regarding the fix ?

The exporting of mask rcnn model should have been resolved in favor of this commit.

Currently the master branch has importing issue, which will be resolved in

Could you try master branch after the above PR get merged?