MXNet DNN Framework

As a newcomer of deep learning, I got to choose what framework(s) is more convenient to learn first.

Now, fast.ai part1 uses theano, part2 uses TF, the lectures are awesome (I’m a lecture 5-1 as of now), and the fact that we use keras makes us fairly framework-independent, but I’m wondering whether TF really is the best choice for a second course in DL, since even using keras one could leverage framework-specific features, and mxnet seems to be superior.

Indeed, I stumbled in that post which left me a bit surprised:

Furthermore, from:

https://www.quora.com/how-is-MxNET-still-surviving-as-a-framework-in-front-of-tensorflow

mxnet seems to be the advanced framework of choice for a production environment, while conceding a bit about flexibility to TF when it comes to frontline reasearch.

I’d like to hear opinions from both fast.ai team and you all.

Actually I find MXNet to be a great framework, because it provides both static and dynamic graphs, making it very very flexible