Theano softmax for tensors?

It seems theano does not having the same capability as tensorflow for softmax. The theano version only allows a matrix input. I have N x T matrices. Each matrix has S x S dimension.

How have people been getting around missing this low level function? Do you do all the sigmoid, reduce sum, and divisions one step at a time?

1 Like

I usually reshape the input, apply the function, and reshape it back.

Thank you, That worked.

Since it seems like you cannot specify the axis as with the softmax function in Tensorflow, reshaping is the work-around, but what I would like to confirm, is the softmax applied to the last axis (-1) then?