Custom operators (RoiPooling) with ONNX

1

I have a model with a custom RoiPooling operator that inherits from torch.autograd.Function. When trying to convert the pytorch model to caffe2, I’m getting the error:

  File "/home/thomasbalestri/PycharmProjects/pytorch-faster_rcnn/tools/../lib/nets/network.py", line 110, in _roi_pool_layer
    return RoIPoolFunction(cfg.POOLING_SIZE, cfg.POOLING_SIZE, 1. / 16.)(bottom, rois)
RuntimeError: Tracing of legacy functions is not supported

I see that there’s a list of supported operators here. Does this mean that there’s no way to export pytorch models with custom operators to caffe2?

2

Hi @Feynman27, right now, we don’t support neither custom operators nor legacy operators.

For custom operators and legacy operators, later we may expose APIs to allows users to define the translation rules.

3

Is there any progress on this since Sep 2017? If I have a custom function with forward and backward passes, how can I export it to onnx?