The ONNX network's output 'output' dimensions should be non-negative

Ask a Question

Question

I trained a CycleGAN and imported it to onnx format. When I import the onnx model to the runtime(Snap Lens Studio), I got the error like this and cannot import it. Resource import for /Users/youjin/Downloads/latest_net_G (1).onnx failed: The ONNX network's output '625' dimensions should be non-negative From this post, I assumed that I got the error cuz the output node does not know its shape. (when I print the output.shape or use torchsummary, it shows the output shape clearly) I used dynamic_axes of torch.onnx.export like the snippet below. When I open it with Netron, it seems to show the right shape, but it still gives the same error on Snap Lens Studio.

I am thinking of using resize at the end of the model now. I’d like to hear any opinion on solving this situation. Even I am not sure I am on the right track.

Further information

• Relevant Area (e.g. model usage, backend, best practices, converters, shape_inference, version_converter, training, test, operators): runtime error

• Is this issue related to a specific model?
  Model name (e.g. mnist): CycleGAN Model opset (e.g. 7): 11

Notes

Any additional information, code snippets.

  net.eval()
  net.cuda()  
  RESOLUTION = 512

  batch_size = 1  
  input_shape = (3, RESOLUTION, RESOLUTION)  # in my case its 512
  export_onnx_file = load_filename[:-4]+".onnx"  
  save_path = os.path.join(self.save_dir, export_onnx_file)

  input_names = ["input"]
  output_names = ["output"]
  random_input = torch.randn(1, 3, RESOLUTION, RESOLUTION, dtype=torch.float32)
  torch.onnx.export(net, random_input, save_path,
                    input_names = input_names, output_names = output_names,
                    opset_version = 11,)
                     dynamic_axes={
                     'output' : {
                       0:'1',
                       1:'3',
                       2:'512',
                       3:'512'
                     }})
  onnx.save(onnx.shape_inference.infer_shapes(onnx.load(save_path)), save_path)

[youjinChung]

I came back with an update. It turns out that Snap Lens Studio has some limitations with onnx models, so it would be safe to use their provided models only. I found the CycleGAN model from Snap Research, and I could import their model without any error.

[youjinChung]

@jcwchen Thanks for checking the model! Now I am assured it is now the model problem, so I can switch my direction. Let me come back if I have any updates or some other onnx-related questions.