Difference between torch.nn.CosineSimilarity and torch.nn.functional.cosine_similarity?

What is the difference between cosine similarity functions torch.nn.CosineSimilarity and torch.nn.functional.cosine_similarity ? Where can we use and where we shouldn’t ? If we are using it inside a model with back-propagation flow, which one can we use ?

Hi Ramraj!

The two are effectively the same and they can be used essentially
interchangeably. In particular, they both support backpropagation
in the same way.

CosineSimilarity is the class / function-object version (and almost
certainly merely wraps a call to cosine_similarity()), while
cosine_similarity() is the function version.

It’s primarily a matter of style which you choose to use. However,
CosineSimilarity is a Module, so you could, for example, place
it in a Sequential, if that were to make sense for your use case.


K. Frank

1 Like