Torch support in ML.NET

[artidoro]

As ML.NET increases it's support for deep learning, we should add support for PyTorch models.
TorchSharp has bindings for LibTorch (PyTorch's C++ frontend) and offers the following functionality:

• Score models that were previously defined and trained in PyTorch's python API
• Define and train models directly in C#

We could have a scoring functionality similar to the TensorFlow scoring. We could also add training of Torch models defined in C#.

[GeorgeS2019]

Now we have the official Microsoft.ML.TorchSharp namespace.

We could close this issue now :-)

[michaelgsharp]

@GeorgeS2019 While we do have the namespace, it currently only has a single TextClassification API. We want to add in support to do more of what is mentioned in this issue like defining TorchSharp models in C# and doing the full training. We are also tracking this elsewhere so we can close this issue, but we don't have all the features yet this issue mentions.

@artidoro good to see you are still around and paying attention to github notifications :)

[artidoro]

@michaelgsharp it's good to see Torch being integrated in ML.NET!
Hope everything is going well :)

[GeorgeS2019]

@michaelgsharp

it currently only has a single TextClassification API.

Based on the Single example, it seems to me that it is doable to port TorchAudio and TorchVision examples from TorchSharp to ML.NET using the same way how the TexTclassification API is set up. This is just a thought, not sure if this is that straight forwards?