Spconv with torchscript


I’m having trouble using torchscript with my model which uses spconv. I need spconv, since I’m working with large sparse matrices. However,

traced_script_module= torch.jit.trace(model, example)

throws the following error:

RuntimeError: Found an unsupported argument type c10::List<at::Tensor> in the JIT tracer.

I’m also getting a lot of transformation errors like:

TracerWarning: Converting a tensor to a NumPy array might cause the trace to be incorrect.

Any Idea what could cause this?