[WIP] Add TorchScript and C++ support

[Mika412]

This is just the TorchScript code. I still need to do the following tasks:

• Refactor the main script
• Fix results bug + add a model results comparison script
• Fork modified pypose library (here)
• Fix TorchScript save method
• Implement C++ sample

[Mika412]

I've added a script based on some of your previous work, to compare the trajectories between the solutions. The TorchScript compiled model gives 1:1 results. but unfortunately, ONNX doesn't. I assume this may be due to the rounding error. Because these values are so small, it may be just rounding them to 0.

[NicolayP]

That is when no acitons are applied to the system right? but also that isn't too much of an issue I think, especially when I look at the order of magnitude in the error ~1e-13. That is for a horizon of 20 steps right? How bad is the divergence when you increase that to 50 and 100?

[Mika412]

I'll try to see if I can check it tonight, as it'll take a while to export.

[Mika412]

Checked 50 and 100. ONNX still gives the same result, probably some torch function is getting incorrectly exported. TorchScript is giving exactly the same result as the original model.

Unfortunately, I couldn't finish the save method before I leave. You can either try fixing it yourself or I can continue working on it when I'm back in August.