Another deep learning framework

Yesterday, Werner Vogels announced that Amazon will throw its weight behind MXNet, a deep learning framework created at Carnegie Mellon. Machine learning will be a major component of re:Invent next week and it sounds like they might be gearing up to announce some tooling around or support for MXNet. They join a host of big tech companies promoting their preferred deep learning tool: Google has TensorFlow, Facebook has Torch, Microsoft has CNTK, Intel has Neon. With so many choices, how do you go about picking the right framework? Here’s how we’re thinking about it.

Hudl R&D uses Python for the vast majority of our work, which means we want Python to be a first-class citizen, ruling out Torch and CNTK. We really love using open source code and pre-trained models to vet network architectures before folding them into our codebase, which makes widespread adoption high on our wish list. TensorFlow definitely wins here. Similarly, good documentation and lots of effective Stack Overflow answers make our lives easier. I recently saw this tweet from someone at Google:

+1 TensorFlow. And while efficiency matters, we value speed of development over compute time and we have enough access to powerful servers that we feel like we can live with slower performance for now.

That said, we’ll be watching Neon and MXNet very closely for the next year or two. In a recent press release, Intel mentioned that they expect deep learning to be 100x faster by the end of the decade. 100x speed improvement would be a game changer and would definitely make us consider switching to Neon. And although we’re happy to do all our work on servers today, I can imagine a world where we’d want to do inference on inexpensive hardware or mobile devices. That would make MXNet more interesting.

Obviously, things are changing fast in this space and we’re not overly eager to write anything in stone. TensorFlow is very effective for our use-case today, but we love the competition among solid open source tools and we’ll be happy to pivot if another framework becomes a better fit.