A big reason is the dynamic computation graphs. Tensorflow fold is supposed to be the response to that, but overall people are finding pytorch to be more flexib

A big reason is the dynamic computation graphs. Tensorflow fold is supposed to be the response to that, but overall people are finding pytorch to be more flexible. (Warning: I'm biased. I compete VERY heavily with anything google including cloud and their deep learning framework.)