[mlthoughts2018]

I would say keras or mxnet for speed and production. PyTorch for research. By this point there are hardly any cases when it’s worth it to descend to lower TensorFlow levels.