question about model export to onnx

I want to use TensorRT to delpoy the model. and I try to export the model to onnx, but I found this error when I use export_onnx function

RuntimeError: Only tuples, lists and Variables supported as JIT inputs/outputs. Dictionaries and strings are also accepted but their usage is not recommended. But got unsupported type numpy.ndarray

and this seems to be because forward function use dict as input param. can I have some non-destructive methods to support export to onnx?

[muzi2045]

pytorch don’t support dict when export model to onnx. so here is something to change in networek input and output, the dict input --> list input here is three part onnx export from OpenPCDet codebase:

[NLCharles]

Thank you for the information. I only tested PartA2 model in pytorch and it’s not far from 10fps. I thought transplanting it to tensorrt could boost its speed.

Could you please tell me which ops are not GPU-friendly, e.g sparse convolution?

Looking forward to your results.

Thank you for your reply. Did your run time include preprocess and post process?

Ops utilized by PointNet++ modules, set abstraction and forward interpolation, involve intensivelogical oprations.

For voxelnet, voxelization of the input costs a lot of time, and sparse convolution must be wrapped into tensorrt plugin.