I know that mxnet was the fastest, highly scalable, highly distributed framework out there. And that was the reason why I switched from tensorflow to mxnet. But after releasing tensorflow 2.0 I am a bit skeptical now. Is mxnet still faster than tensorflow? Any thoughts would be appreciated.
Hi @mouryarishik, I haven’t seen any indications that TensorFlow 2.0 will be any faster than before, so I don’t think the situation has changed in this regard. It’s mostly frontend API changes as I understand.
I’ve seen the keynote and there they are saying that TF 2.0 on gpu is almost 1.8x faster than previous version.
One thing that I liked about TF 2.0 is that it allows us to code control flow by using native python if-else, for loop, while loop in dynamic way, and it would automatically convert your code into static graph if needed. Whereas in mxnet you must have to use mx.nd.contrib.foreach, .while_loop, .cond in order to convert your code from dynamic to static.
Correct me if I’m wrong anywhere.
Oh, that’s interesting! Are you talking about the Autograph API here? Yes, with MXNet Gluon you can use native control flow operators in imperative mode, but for hyridization you then use the control flow operators.
And I agree that it would be good to see some benchmarks with TensorFlow 2.0 when it becomes more stable: it’s currently in alpha as I understand.
Hey check this out
It’s new updated quick and dirty benchmark I’ve done again. And mxnet is indeed faster than tensorflow
Thanks a lot! Do you know any other benchmark on different models?
Check this out, here is a benchmark recently done by Nvidia:
And welcome to the community.