CNTK integrated into Keras

June 13, 2017

(This article was originally published at SAS Programming for Data Mining, and syndicated at StatsBlogs.)

Keras is a very convenient tool to build your deep learning model from scratch, it is so easy to use that it almost becomes the de-facto deep learning modeling framework in Kaggle competition.

Keras used to support only TensorFlow and Theano.  CNTK, now rebranding as Cognitive Toolkit in full name, from Microsoft is joining as a new backend choice, but you need to download Microsoft private fork of Keras at this moment while Microsoft is working with Keras author to fully integrate CNTK as its backend.

CNTK is a very good deep learning tool: 1. it is super fast especially for RNN type of model; 2. it scales well across multiple GPUs Here is our own speed comparsion (number in sec) on Windows platform using NVidia Titan Xp, YMMV though:
If you want to give CNTK a try, you can follow the installation instruction @ here. To install Keras that supports CNTK, follow the instructions @here.

Please comment on the article here: SAS Programming for Data Mining