[Bug or Requirement?] ONNX unpacked the tuple of parameters of the model

[thlinh]

If you have a question or would like help and support, please ask at our
forums.

If you are submitting a feature request, please preface the title with [feature request].
If you are submitting a bug report, please fill in the following details.

Issue description

Hi,
I have a problem when converting a PyTorch model into ONNX. My model expects its input is a tuple of tensors (one input only). But when running torch.onnx.export(), the input is unpacked and passed into the model as a number of tensors, which generates error (since only 1 parameter is expected). I have a simplified version shown down here:

Code example

import torch
import torch.nn as nn

class sim_test(nn.Module):
    def __init__(self, args):
        super(sim_test, self).__init__()
        print(args)

    def forward(self, inp_tuple):
        x = inp_tuple[0]
        print(x.type())
        print(x.size())
        y = inp_tuple[1]
        z = inp_tuple[2]
        return x + y + z

def main():
    model = sim_test("Test me")
    inp_x = torch.FloatTensor(1, 5).uniform_()
    inp_y = torch.FloatTensor(1, 5).uniform_()
    inp_z = torch.FloatTensor(1, 5).uniform_()
    total_inp = (inp_x,inp_y,inp_z)
    result = model(total_inp)
    print ("Result is: ", result)

    torch.onnx.export(model, total_inp, "simple_test.onnx", verbose=True)

if __name__ == '__main__':
    main()

The result screen is:

torch.FloatTensor
torch.Size([1, 5])
Result is:  tensor([[1.2815, 1.8511, 0.9751, 1.1096, 1.4900]])
Traceback (most recent call last):
  File "simple_test.py", line 30, in <module>
    main()
  File "simple_test.py", line 27, in main
    torch.onnx.export(model, total_inp, "simple_test.onnx", verbose=True)
  File "/home/h00472437/.conda/envs/Py36/lib/python3.6/site-packages/torch/onnx/__init__.py", line 26, in export
    return utils.export(*args, **kwargs)
  File "/home/h00472437/.conda/envs/Py36/lib/python3.6/site-packages/torch/onnx/utils.py", line 94, in export
    operator_export_type=operator_export_type)
  File "/home/h00472437/.conda/envs/Py36/lib/python3.6/site-packages/torch/onnx/utils.py", line 227, in _export
    example_outputs, propagate)
  File "/home/h00472437/.conda/envs/Py36/lib/python3.6/site-packages/torch/onnx/utils.py", line 178, in _model_to_graph
    graph, torch_out = _trace_and_get_graph_from_model(model, args, training)
  File "/home/h00472437/.conda/envs/Py36/lib/python3.6/site-packages/torch/onnx/utils.py", line 145, in _trace_and_get_graph_from_model
    trace, torch_out = torch.jit.get_trace_graph(model, args)
  File "/home/h00472437/.conda/envs/Py36/lib/python3.6/site-packages/torch/jit/__init__.py", line 77, in get_trace_graph
    return LegacyTracedModule(f)(*args, **kwargs)
  File "/home/h00472437/.conda/envs/Py36/lib/python3.6/site-packages/torch/nn/modules/module.py", line 477, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/h00472437/.conda/envs/Py36/lib/python3.6/site-packages/torch/jit/__init__.py", line 109, in forward
    out = self.inner(*trace_inputs)
  File "/home/h00472437/.conda/envs/Py36/lib/python3.6/site-packages/torch/nn/modules/module.py", line 475, in __call__
    result = self._slow_forward(*input, **kwargs)
  File "/home/h00472437/.conda/envs/Py36/lib/python3.6/site-packages/torch/nn/modules/module.py", line 465, in _slow_forward
    result = self.forward(*input, **kwargs)
TypeError: forward() takes 2 positional arguments but 4 were given

The problem can be now by-passed by changing the definition of forward() to:

def forward(self, *inp_tuple):

and call it as:

result = model(*total_inp)

I wonder if this is a requirement that we cannot use directly tuple as input into a model?
Thank you very much for your coming helps! 😃

System Info

Collecting environment information...
PyTorch version: 0.4.1.post2
Is debug build: No
CUDA used to build PyTorch: 9.0.176

OS: Ubuntu 16.04.5 LTS
GCC version: (Ubuntu 5.5.0-12ubuntu1~16.04) 5.5.0 20171010
CMake version: version 3.5.1

Python version: 3.6
Is CUDA available: Yes
CUDA runtime version: 7.5.17
GPU models and configuration: GPU 0: Tesla K40c
Nvidia driver version: 396.51
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.7.1.4
/usr/local/cuda-9.0/targets/x86_64-linux/lib/libcudnn.so.7.1.1
/usr/local/cuda-9.0/targets/x86_64-linux/lib/libcudnn_static.a
/usr/local/cuda-9.2/targets/x86_64-linux/lib/libcudnn.so.7.1.4
/usr/local/cuda-9.2/targets/x86_64-linux/lib/libcudnn_static.a

Versions of relevant libraries:
[pip3] numpy (1.15.1)
[pip3] torch (0.4.1.post2)
[pip3] torchsummary (1.4)
[pip3] torchtext (0.2.3)
[pip3] torchvision (0.2.1)
[conda] pytorch                   0.4.1           py36_py35_py27__9.0.176_7.1.2_2    pytorch
[conda] torchsummary              1.4                       <pip>
[conda] torchtext                 0.2.3                     <pip>
[conda] torchvision               0.2.1                    py36_1    pytorch

[houseroad]

Fix: passing total_inp as a tuple, since it will be unpacked.
torch.onnx.export(model, (total_inp,), "simple_test.onnx", verbose=True)

[thlinh]

@houseroad : Is the tuple, as an input parameter, always unpacked by ONNX when calling a function?