Weird behavior of `CosineSimilarityMetric` when used with tensors of shape [d]

[ValerianRey]

🐛 Bug?

According to the documentation, CosineSimilarityMetric requires tensors of shape [N, d], with N the batch size and d the dimension of the vectors.

Using it with vectors of shape [d] does not raise any error, and the call to compute behaves weirdly and gives confusing results. I'm not sure whether this is the expected behavior, a bug, or if this usage is simply unintended (and then maybe an error could have been raised).

To Reproduce

>>> from torchmetrics import CosineSimilarity
>>> from torch import tensor
>>> 
>>> cosine_similarity = CosineSimilarity(reduction="mean")
>>> a = tensor([1., 1., 1.])
>>> b = tensor([100., 100., 100.])
>>> cosine_similarity(a, b)
tensor(1.)  # a and b have the same direction, so this is normal.
>>> cosine_similarity(b, a)  
tensor(1.)  # same for b and a.
>>> cosine_similarity.compute()
tensor(0.0200)  # I would expect this to be 1 too (the average of the 2 previous calls).

The obtained result (0.02) is actually the cosine similarity between [1, 1, 1, 100, 100, 100] and [100, 100, 100, 1, 1, 1]. I would have expected to get instead the average between the cosine similarity of [1, 1, 1] and [100, 100, 100] and the cosine similarity of [100, 100, 100] and [1, 1, 1], which is 1.

If instead we use it as the documentation says, with tensors of shape [N, d], we get different results:

>>> from torchmetrics import CosineSimilarity
>>> from torch import tensor
>>> 
>>> cosine_similarity = CosineSimilarity(reduction="mean")
>>> a = tensor([[1., 1., 1.]])  # tensor of shape [1, 3] instead of [3]
>>> b = tensor([[100., 100., 100.]])  # tensor of shape [1, 3] instead of [3]
>>> cosine_similarity(a, b)
tensor(1.)
>>> cosine_similarity(b, a)
tensor(1.)
>>> cosine_similarity.compute()
tensor(1.)  # 1 instead of 0.02

Environment:

• TorchMetrics 1.2.0
• Python 3.10.10
• torch 2.1.1
• Ubuntu 20.04.6 LTS

Additional context