Could I convert lightning module to onnx? Thanks!

[stevenluzheng123456]

🚀 Feature

pytorch Lightning works very good, but I cannot find any comments or examples to guide my convert to onnx from a pretrained lightning model, doesn't lightning module only use for researching purpose, without support of onnx cross platform?

[github-actions]

Hi! thanks for your contribution!, great first issue!

[Borda]

good point, @PyTorchLightning/core-contributors thoughts?

[SkafteNicki]

From the docs https://pytorch.org/docs/stable/onnx.html it seems easy enough. We would require users to define the self.example_array attribute, but that's it.

Do we already have a method for saving the model that can be extended?

[williamFalcon]

a lightningModule is just a nn.Module.

Wouldn't this just work?

# Input to the model
x = torch.randn(batch_size, 1, 224, 224, requires_grad=True)
torch_out = torch_model(x)

# Export the model
torch.onnx.export(torch_model,               # model being run
                  x,                         # model input (or a tuple for multiple inputs)
                  "super_resolution.onnx",   # where to save the model (can be a file or 

But we should just make this automatic...
@stevenluzheng123456 want to submit a PR for:

model = LitModel(...)
model.to_onnx(x)

[lezwon]

@williamFalcon I could take up this issue.
Was wondering if we could automatically infer the input via the dataloader, instead of passing it to to_onnx. Something along the lines of:

batch = next(iter(model.train_dataloader()))
input_data = batch[0]
torch.onnx.export(model, input_data, file_path)

[Borda]

@lezwon from which PT version is there this option?

[lezwon]

@Borda Going by the docs, I think ONNX export is supported since 0.3.0 in PyTorch.

[awaelchli]

Could one also use LightningModules example_input_array for this?

[lezwon]

@awaelchli Didn't know about example_input_array. We could use that if it's available. Else we could take the dataloader approach. We could also provide an optional parameter to pass in an Input Tensor.

[awaelchli]

Yes so far the example_input_array is used for the summary printing input and output shapes. It is optional but if the user wants the full summary they need to define it. The reason why the user needs to define it manually is

• module forward may have different input types/shapes than training_step (what dataloader serves).
• dataloader definition is optional in a LightningModule, since these can be passed also to fit/test

To me it looks like exporting to onnx is very similar to the model summary in this regard.

[lezwon]

It does seem very similar. So the way I see it, the example_input_array approach might be better as the forward method might have an input with a different shape. Maybe we could also allow an optional parameter in case example_input_array is not present?
So the approach would be:

1. Check if optional input tensor provided
2. If not provided look for example_input_array
3. If not present throw error.

model.to_onnx(file_path: str, input: Optional[Tensor])

[awaelchli]

I like it! 👍

[lezwon]

Cool. I'll get working on it 😊

[justusschock]

@lezwon Can you please make sure to allow additional kwargs that are passed to the onnx-export? These would come handy for customisation :)

[lezwon]

Will do :)