The cosine similarities

[ypxie]

z, _ = theano.map(step, sequences=[x, y]). Should x be in the non_sequence?

[tristandeleu]

The map operates over the examples in the mini batch, so x in cosine_similarity is effectively a matrix, and y is a tensor3 (there's a mistake in the documentation, thank you for pointing that out!). This is clearly not the best way to loop over the different batches, but I didn't find a cleaner way to directly work on the matrix/tensor3 version unfortunately.