Conversion to/from ONNX models

[tonieichelkraut]

I would really love to work with this library. I have an application that runs onnx-models. Is there a way in diffsharp to convert models to/from the onnx-format? That would really make my day! Or how could I accomplish training models using Diffsharp and running onnx-models in production?

[dsyme]

@tonieichelkraut Realistically you or someone else would have to implement and contribute this. The way to do it would be the same way PyTorch does - write a backend that captures the trace of model construction at a primitive level and output that as an ONNX model.