Could you help me understanding the Similarity functions in DIET Classification?
In the Docs, I found 3 functions which are ‘auto’(used by default), ‘cosine’, and ‘inner’.
Cosine is clear. But the two others are not.
I hope you might be able to give me some insights on this.
Hello,
The inner product is a generalization of the dot product, but in the Euclidean space Rn (the real n-space ) the inner product is given by the dot product.
About the “auto” label I think that it depends on the loss function used. When we use a softmax loss, the similarity_type is the inner product, and if use the margin loss the similarity_type is the cosine similarity.