I’ve been doing a lot of work with CNTK lately (and for the most part I quite like it compared to TensorFlow), but there are two things in particular that need

I’ve been doing a lot of work with CNTK lately (and for the most part I quite like it compared to TensorFlow), but there are two things in particular that need to be addressed before I think it can gain wider traction: support for batch normalization on CPUs (right now bn is GPU-only, which means a lot of Microsoft’s examples only run on Windows/Linux machines with GPUs, not something everyone has immediate access to always), and macOS support (even CPU-only support). I think those things are table stakes for any deep learning framework. That said, it’s great to see the author’s work here - CNTK documentation can be a bit hard to come by.