ONNX Export failed on ATen operator var (KeyError: 'var')

Hi all,

I’m trying to convert a PyTorch model to ONNX with torch.onnx.export, but the operation fails upon trying the ‘var’ operator (symbolic_opset9). I have done some reading and found mean variance normalization (mvn) is supported, but I could not find anything about var alone.

The issue lies when torch.onnx.export is called, with the following trace:

/opt/anaconda3/lib/python3.7/site-packages/torch/onnx/utils.py:715: UserWarning: ONNX export failed on ATen operator var because torch.onnx.symbolic_opset9.var does not exist

Traceback (most recent call last):
File “onnx_conversion_script.py”, line 74, in
torch.onnx.export(model, dummy_input, model_name + “.onnx”, verbose=True, input_names = input_name, output_names=output_name)
File “/opt/anaconda3/lib/python3.7/site-packages/torch/onnx/init.py”, line 158, in export
custom_opsets, enable_onnx_checker)
File “/opt/anaconda3/lib/python3.7/site-packages/torch/onnx/utils.py”, line 68, in export
custom_opsets=custom_opsets, enable_onnx_checker=enable_onnx_checker)
File “/opt/anaconda3/lib/python3.7/site-packages/torch/onnx/utils.py”, line 469, in _export
fixed_batch_size=fixed_batch_size)
File “/opt/anaconda3/lib/python3.7/site-packages/torch/onnx/utils.py”, line 338, in _model_to_graph
fixed_batch_size=fixed_batch_size, params_dict=params_dict)
File “/opt/anaconda3/lib/python3.7/site-packages/torch/onnx/utils.py”, line 153, in _optimize_graph
graph = torch._C._jit_pass_onnx(graph, operator_export_type)
File “/opt/anaconda3/lib/python3.7/site-packages/torch/onnx/init.py”, line 189, in _run_symbolic_function
return utils._run_symbolic_function(*args, **kwargs)
File “/opt/anaconda3/lib/python3.7/site-packages/torch/onnx/utils.py”, line 716, in _run_symbolic_function
op_fn = sym_registry.get_registered_op(op_name, ‘’, opset_version)
File “/opt/anaconda3/lib/python3.7/site-packages/torch/onnx/symbolic_registry.py”, line 94, in get_registered_op
return _registry[(domain, version)][opname]
KeyError: ‘var’

Is a custom operator necessary? Has anyone else run into this? I have tried both the latest PyTorch as well as the nightly build with no success.

Thanks.

Variance doesn’t seems to be an ONNX operator. But hopefully, it can be expressed as a combination of simpler operations.

  • One way to “bypass” it is to remove your call to “var” and replace by its formula.
  • A better approach is to add to torch.onnx the var export (I could do it if I get some free time this evening).
  • Ideally, if its usage is common, it would be interesting to open an issue on the ONNX Repo to ask for a Variance opcode that could be added to the standard.

Do you know how many models use this operator? I never had to use it before personally. :sweat_smile:


PS: The variance can be expressed as {\displaystyle \sigma ^{2}={\overline {(x^{2})}}-{\bar {x}}^{2}=\displaystyle {\frac {\sum {i=1}^{N}x{i}^{2}-(\sum {i=1}^{N}x{i})^{2}/N}{N}}.!}

1 Like

@jaolan I think you should open a feature request on https://github.com/pytorch/pytorch/issues

Thanks for your reply!

Adding var to the export is a good idea :grin:. This is the first time I’ve seen it in a model I am exporting as well, but it seems like an operator that could be useful to others - I will open a feature request in PyTorch.

Good idea! I will open an issue in PyTorch requesting the Var operator.