[JIT] [Mobile] Op may not exist in TorchScript

Hello everyone!

Once again I’ve run into something I don’t understand, need your help.

I’ve compiled speed_benchmark_torch for mobile to test my TorchScript RCNN model.

The model is from detectron2. I’ve slightly varied caffe2 conversion of RCNN to use torch.jit.trace instead torch.onnx.export https://github.com/facebookresearch/detectron2/blob/master/tools/deploy/caffe2_converter.py
That gives a valid model to run on PC but on Mobile there is an issue.

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.
Serialized   File "code/__torch__/detectron2/export/c10.py", line 22
    scores = torch.detach(_6)
    bbox_deltas = torch.detach(_7)
    _16, _17 = ops._caffe2.GenerateProposals(scores, bbox_deltas, im_info, _4, 0.25, 1000, 100, 0.69999999999999996, 0., True, -180, 180, 1., False, None)
               ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ <--- HERE
    scores0 = torch.detach(_8)
    bbox_deltas0 = torch.detach(_9)

The question is possible to un TorchScript model with caffe2 ops or we should use old speed_benchmark for caffe2 models?
Why there are caffe2 related options in ./speed_benchmark_torch --help?

Hi @zetyquickly, I was curios if you every found a solution to your problem. I am currently running into basically the same error trying to deploy a model to ios. Any tips would be greatly appreciated thanks!

Hello, @Kyle_Armstrong

I’m afraid that you won’t be able to use PyTorch backend to run Caffe2 operations on mobile. If you are interested in deploy of Caffe2 models you should build a Caffe2 .pb model files and then run it with Caffe2 backend on mobile, e.g. using speed_benchmark (not speed_benchmark_torch)
To run models using only PyTorch backend your TorchScript model must contain only PyTorch operators and only supported on with mobile PyTorch (e.g. you’re unable to run torchvision::nms op on mobile, at least that was the case recently)

1 Like