Support defining NDArray-like types for NumPy-ish frameworks?

[EricCousineau-TRI]

I'd like to define something like Tensor for use with pytorch (relates #27).

Some random notes:

• This would loosely enable any frameworks that offer array-like objects (e.g. torch, tensorflow, open3d (code)), generally something that offers .shape and .dtype that behave similar to NumPy's
• Ideally, some of this work could also permit ndarray subclasses (docs), and handle the whole covariance shindig mentioned in PEP 484 (docs)

[EricCousineau-TRI]

I, uh, couldn't figure out how to make PyTorch easily play well with _NDArray and all the generics, so I went ahead and wrote a poor-man's duplicate. I've hashed out a simple version of this here:
https://github.com/EricCousineau-TRI/repro/blob/4a5c71162bf2e3c0afc4d08c5d757aa0841000c4/python/nptyping_ish/test/typing_test.py

[EricCousineau-TRI]

FWIW I've updated the prototype to allow working with dataclasses and as a (semi-annoying) stand-in for default_collate:
https://github.com/EricCousineau-TRI/repro/blob/7a08250e55899dd47fb77ac6524331cd24d86e59/python/nptyping_ish/typing_/test/typing_test.py#L293-L427

[ramonhagenaars]

Good idea to support similar frameworks. I was thinking of adding support for pandas, tensorflow and pytorch would be nice additions too.

The other frameworks are to become optional for nptyping, meaning that you have to explicitly specify them when installing:

pip install nptyping[pandas]
pip install nptyping[tensorflow]
...
pip install nptyping[complete]

I think I'll start designing an addition for Pandas first. Then I could use help with other frameworks (I don't know open3d for example).

[EricCousineau-TRI]

Sounds great!

And yeah, for open3d, I think it'd basically be a restricted subset of NumPy functionality, so nothing fancy. (Just may have device-like semantics like PyTorch / TensorFlow, but no computation graph thingers for reverse autodiff going on.)

[zplizzi]

Would love to see this for torch :)

[dxm447]

Would love to see it for cupy too

[dxm447]

cupy.typing has been available since this year, and is a like for like replacement of np.typing