CplxConv1d can be exported to ONNX but cannot be inferred by ONNXRUNTIME

[pfeatherstone]

The following example code shows that cplxmodule.nn.CplxConv1d can be exported but cannot be run.

    class Net(nn.Module):
        def __init__(self):
            super(Net, self).__init__()
            self.net = cplxmodule.nn.CplxConv1d(1,1,3,padding=1)

        def forward(self, x):
            x = cplxmodule.Cplx(x[..., 0], x[..., 1])
            x = self.net(x)
            x = torch.stack([x.real, x.imag], dim=-1)
            return x

    model = Net().eval()

    input = torch.randn(1, 1, 1024, 2)
    out = model(input)

    torch.onnx.export(model,
                      (input,),
                      "file.onnx",
                      opset_version=12,
                      input_names=['in'],
                      output_names=['out'])

    print("exported")

    import onnxruntime
    ort_session = onnxruntime.InferenceSession("file.onnx")
    ort_inputs = {ort_session.get_inputs()[0].name: input.numpy()}
    ort_outs = ort_session.run(None, ort_inputs)
    assert len(ort_outs) == 1, "bad number of outputs"
    np.testing.assert_allclose(out.detach().cpu(), ort_outs[0], rtol=1e-04, atol=1e-05)
    print("done")

[pfeatherstone]

I get error:

[E:onnxruntime:, sequential_executor.cc:318 Execute] Non-zero status code returned while running Slice node. Name:'Slice_11' Status Message: /onnxruntime_src/onnxruntime/core/providers/cpu/tensor/slice.cc:221 static void onnxruntime::SliceBase::FillVectorsFromInput(const onnxruntime::Tensor&, const onnxruntime::Tensor&, const onnxruntime::Tensor*, const onnxruntime::Tensor*, std::vector<long int>&, std::vector<long int>&, std::vector<long int>&, std::vector<long int>&) ends_tensor.Shape().NumDimensions() == 1 was false. Ends must be a 1-D array

[pfeatherstone]

I highly doubt this is an onnxruntime bug. But then again, the debug statements give no useful hints.

[ivannz]

Thank you for your catch @pfeatherstone . the issue has been fixed in pr #14