Adds example of inferring dimensions.

[zimonitrome]

This feature is a nugget that might be obvious to some but is hidden in any documentation that I am able to find.

My edit might be on the nose of the readme but I otherwise suggest to include some other example to showcase inference of dimensions, e.g.:

image_sample = torch.rand([3, 256, 256])
image_batch = torch.rand([10, 3, 256, 256])

def make_grayscale(t):
    return reduce(t, "... c h w -> ... h w", "mean")

image_sample_gray = make_grayscale(image_sample)
image_batch_gray = make_grayscale(image_batch)

print(image_sample_gray.shape)
print(image_batch_gray.shape)

> [256, 256]
> [10, 256, 256]

[vmichals]

I assumed that this existed and should've guessed that the correct syntax is '...', but didn't! Luckily I saw your PR! Thanks!
I agree this is a very useful example for the documentation!

[charlesjhill]

^ I found one example of using ... as a catch-all in the docs, at https://einops.rocks/pytorch-examples.html:

if combine == 'add':
    # ShuffleUnit Figure 2b
    self.left = Rearrange('...->...') # identity
    depthwise_stride = 1

An example like in the PR or something more explicit out ala

data_bthwc = np.random.random([2, 10, 100, 110, 3])
data_btchw = rearrange(data_bthwc, "... h w c -> ... c h w")
data.shape  # (2, 10, 3, 100, 110)

would have saved me some time on a recent project.

[arogozhnikov]

all PRs were auto-closed because of switching master -> main, hopefully I'll have some time to update documentation, and include ellipsis with enough context to prevent lazy usage.