Remove NumPy as a hard dependency

[dgasmith]

Description

This PR removes NumPy from opt_einsum as a runtime dependency. The change has little effect on the opt_einsum code base but significantly effects the testing infrastructure. Several testing strategies are currently employed to help understand the most effective one. Currently, these are:

• Move all NumPy functions to a testing.py module with pytest.skip
• Various pytest.skip functions wrapped in decorators
• Local NumPy imports.

Feedback welcome!

Todos

Notable points that this PR has either accomplished or will accomplish.

• Remove unused custom tensordot code
• Fix ssa_to_linear with a proper implementation
• possibly_convert_to_numpy is used throughout, primarily to give objects a shape attribute if they do not have one. Setting attributes on root types is difficult and would require subclassing int/float/etc to provide a shape. I'm kicking around possible solutions here.
• Fully differentiating GenericArrayType (without a shape attribute) compared to ArrayType (with a shape attribute) throughout the code base.
• We need a torch-based only test case.

Questions

• Question1

Status

• Ready to go

[jcmgray]

LGTM so far! Do you understand the test failures, is that just an env thing?

[dgasmith]

I didn't get to every test yet. I'll hack a bit this weekend.

[dgasmith]

@janeyx99 Is this still of interest to Torch? I think we're relatively close on removing that last of the NumPy dependancies.

[janeyx99]

Yes! Removing the last dependencies would be awesome!

[dgasmith]

Apologies- I started a new job. This is in my mind and I hope to hack a little tomorrow night!

[dgasmith]

@janeyx99 I'm looking to push another release shortly, is this still a requested feature?

[janeyx99]

@dgasmith yes, torch still doesn't have a hard dependency on numpy (our tests do) so this is still desirable! Sorry it took a week to get back to you--this fell in my personal inbox.

[dgasmith]

I'll get in #232 first and then finalize this PR. It will be quite tidy to have no formal dependancies.

[codecov]

Codecov Report

Attention: Patch coverage is 92.68293% with 12 lines in your changes missing coverage. Please review.

Project coverage is 93.77%. Comparing base (1a984b7) to head (8211264).
Report is 5 commits behind head on master.

Additional details and impacted files

[dgasmith]

@jcmgray I think this is getting much closer. I updated the outstanding items in the OP.

[dgasmith]

@jcmgray and @janeyx99 This is ready for review, but likely has a few bugs to still iron out. In particular @janeyx99 is there any tests you would recommend for PyTorch only envs?

[jcmgray]

Looks good to me!

[janeyx99]

@jcmgray and @janeyx99 This is ready for review, but likely has a few bugs to still iron out. In particular @janeyx99 is there any tests you would recommend for PyTorch only envs?

Thanks for furthering this effort! The main reason this would be nice for PyTorch is so that we allow opt-einsum to be a more accessible dependency vs one the user has to specify when pip installing torch.

To that end, we have a few test commands written in bash to ensure that we don't accidentally depend on numpy:
https://github.com/pytorch/pytorch/blob/ca5d13c67287efe78a09dcf9b06f7e4fba898689/.ci/pytorch/test.sh#L742-L752

I don't think it's crucial for opt-einsum to have similar tests, because we will test these as a downstream library, but I'm mentioning in case you were looking for similar sanity checks.

[dgasmith]

@janeyx99 I do have a torch-only env explicitly for testing non-numpy evaluations. I like adding the check to ensure that NumPy is not present.