Sasha Rush

This looks great. the isalphanum stuff could go right in if you wanted to send a separate PR. One minor thing: I want to move all the asserts to RuntimeError's....

Yes, I mostly agree with this, I think we are on the same page. However, I would 1) like indices to be a named tensor too not a list, 2)...

Wow, so cool! Did you find that the current named tensors proposal matches the blog post? There are some new aspects that are pretty cool and different. I'm also not...

Oh, actually, there is one issue. As far as I know the namedtensor currently in pytorch is deprecated, and they have been experimenting with a new version here: https://github.com/facebookresearch/torchdim ....

Good question. I got the sense from talking with the torch team that the current implementation was kind of a dead end. The new one is pretty interesting and I'd...

Another random idea would be to write this in Jax. Their implementation is really cool and very close to the original. https://jax.readthedocs.io/en/latest/notebooks/xmap_tutorial.html

So I think it would be fun to do TorchDim then. Not sure if everything translates over but it feels like it might be a bit more future oriented. But...

This is neat, but I don't think I will add it. I think this type of feature should be done with variables and not shorter strings.

Thanks for the suggestions, this is helpful and I hear the two different approaches. I agree the dict construction approach is a bit heavyweight. particularly in loops. I need to...

Could you post a regular tensor? Just curious.