Including LibTorchvision on Android

Area Select

react-native-pytorch-core (core package)

Description

Hello, thanks for your contributions.

I encountered a error on react native android to loading custom ptl model. I think this error can solve if i can include Libtorchvision from relavated issue on this: https://github.com/facebookresearch/playtorch/issues/64

How can i include it in my android application?

Please provide me any leads. Thank you.

Error: {"message": "Following ops cannot be found. Please check if the operator library is included in the build. If built with selected ops, check if these ops are in the list. If you are a Meta employee, please see fburl.com/missing_ops for a fix. Or post it in https://discuss.pytorch.org/torchvision::nms () Exception raised from initialize_operators at /data/users/atalman/pytorch/torch/csrc/jit/mobile/function.cpp:89 (most recent call first): (no backtrace available)"}

[raedle]

Hi @nh9k, that’s correct. The nms (non-maximum suppression) op is part of the Torchvision ops. The react-native-pytorch-core doesn’t include the Torchvision ops, but they can be added to a project, if needed.

I created a repo for the D2Go model, which also requires Torchvision ops to run inference. I created a single commit for the changes that add Torchvision ops to Android and iOS. Have a look at this commit in the raedle/D2GoExample repo!

Note: for iOS it uses the LibTorchvision-DONOTUSE CocoaPod because the Torchvision ops 0.13.0 that are compatible with PyTorch 1.12 have not been released officially yet. They are coming soon, and in the meantime you can use the LibTorchvision-DONOTUSE, which are functionally equivalent to the upcoming release!

Hope that helps, and please report back if you are successful! Any feedback to the PlayTorch is helpful for us to learn 😃

Thank you!

[raedle]

@nh9k, from the logcat logs it looks like that the shared object libraries aren’t packaged correctly.

But generally, there are two parts that are important to add torchvision_ops to your app. See how this is done in the PlayTorch app:

1. Add the torchvision_ops to the app’s build.gradle (make sure the versions for torchvision_ops and pytorch_jni_lite are compatible)
2. Load the torchvision_ops shared object library after loading the pytorch_jni_lite shared object library (see MainApplication.java)