OnnxRuntime C# library - add support for Sequence and Map model output types

[jignparm]

Add support in C# OnnxRuntime Library for sequences and maps (complex types). Tensors are the only supported type currently.

[snnn]

That requires the C API supports complex types, which should be a huge work. Because sequences and maps are not the only non-tensor types. Instead of doing them one by one, we should have a clear roadmap first.

[jignparm]

There's a model that a user converted to Onnx (see dotnet/machinelearning#1323), which produces a Sequence<Map<int,float>> as an output. It seems like a valid ONNX model, and therefore should able to run inference on it. Is there a non C API which is running this inference currently?

[snnn]

The python binding is using the C++ API, but C# is using the C API, because it is not possible to access C++ functions through P/Invoke.

[derdanielb]

Any progress to this? @jignparm

Loading a model which outputs Sequence<Map<int,float>> does not work using the latest C# onnxruntime nuget package (v 0.2.1).
However loading the model using python onnxruntime works fine.

I could really use an implementation of this in the C# onnxruntime.

[jignparm]

This functionality is in progress (i.e. the C API supports it now), and surfacing it in C# is in progress. Should be available next week.

[jignparm]

@derdanielb - please look at the example below, for a model which produces seq<map<int,float>> type.

https://github.com/Microsoft/onnxruntime/blob/3a2f6c69646df66fe82c919cd956a2034a418a0e/csharp/test/Microsoft.ML.OnnxRuntime.Tests/InferenceTest.cs#L528