[Feature suggestion] fixup/support anonymous axes in parse_shape

[rehno-lindeque]

It would be handy to support the second example below:

>>> parse_shape(torch.tensor([[0]]), "b n")
{'b': 1, 'n': 1}

>>> parse_shape(torch.tensor([[0]]), "b 1")
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/nix/store/2x7c1xn335dvgia34gz2dqz6vy3a9pj0-python3-3.11.6-env/lib/python3.11/site-packages/einops/einops.py", line 690, in parse_shape
    for (axis_name,), axis_length in zip(composition, shape):  # type: ignore
        ^^^^^^^^^^^^
ValueError: not enough values to unpack (expected 1, got 0)

Example use case:

def assert_shapes(*tensor_patterns, **dim_sizes):
    # Collect all shapes, including explicit dimension sizes
    shapes = [
        dict(**dim_sizes),
        *(parse_shape(tensor, pattern) for tensor, pattern in tensor_patterns),
    ]

    # Check for consistency in dimensions across all tensors
    for dim_name in set(dim for shape in shapes for dim in shape):
        sizes = set(shape.get(dim_name, None) for shape in shapes)
        sizes.discard(None)
        assert len(sizes) <= 1, f"Inconsistent size for dimension '{dim_name}': {sizes}"

tensor1 = torch.randn(3, 1, 64)
tensor2 = torch.randn(3, 2, 64)

# I want:
assert_shapes(
  (tensor1, 'batch 1 h'),
  (tensor2, 'batch 2 h'),
  h=64
)

# Instead of:
assert_shapes(
  (tensor1, 'batch c1 h'),
  (tensor2, 'batch c2 h'),
  h=64,
  c1=1,
  c2=2
)

I imagine it would have similar behavior to the ellipsis / underscore syntax:

>>> parse_shape(torch.tensor([[[1,2,3]]]), "b _ k")
{'b': 1, 'k': 3}

Thanks!

[arogozhnikov]

Makes sense, please try current master:

pip install git+https://github.com/arogozhnikov/einops.git

PS seems einops in nix-able now 😉

[rehno-lindeque]

Thank you, the ergonomics of einops just keeps getting better!

PS seems einops in nix-able now 😉

Indeed, since 0.3.2 actually 😄