Create generic einops.einop operation to do any of rearrange/reduce/repeat

[MilesCranmer]

As mentioned in #84, the einop operation can be inferred based on pattern alone. This PR adds a function to do this: einops.einop. For example:

Reduction:

import numpy as np
from einops import einop

x = np.random.randn(100, 5, 3)

einop(x, 'i j k -> i j', reduction='sum').shape
>>> (100, 5)

Rearrange:

einop(x, 'i j k -> k i j').shape
>>> (3, 100, 5)

Repeat:

einop(x, 'i j k -> i j k l', l=10).shape
(100, 5, 3, 10)

There are also assertions to catch bugs as mentioned in #84. I also added some tests for these operations: test12 through test14.

Interested to hear what you think.
Cheers!
Miles

[genghisun]

Such a great feature! 👍 Any update on this?

[MilesCranmer]

^See #84 for discussion. Currently, the maintainer is against adding this

[cgarciae]

I've been waiting for this PR to land for a while, in the meantime I've extracted this functionality into the einop package.

[FrancescoSaverioZuppichini]

guys?

[MilesCranmer]

I wish we could, but nothing we can do unfortunately. If you have a compelling argument to convince the maintainer, you can add it to #84. In the meantime you can use @cgarciae’s package

[UFO-101]

@MilesCranmer Thoughts on doing a similar thing to einsum? #84 (comment)