Comments on: Deep Learning Frameworks: A Survey of TensorFlow, Torch, Theano, Caffe, Neon, and the IBM Machine Learning Stack

By: Anupam Singh

Anupam Singh — Tue, 03 Jan 2017 07:20:58 +0000

I worked at nVidia few years back and was involved in rolling out its general purpose computing platform, mainly working on low level software modules. The breadth of John’s experience is amazing. I too was very interested in Physics before college but once there I got fascinated by Computer Science. Nice Post!

By: Eliot Eshelman

Eliot Eshelman — Mon, 07 Nov 2016 20:21:32 +0000

In reply to Arif. Yes. The Microsoft Cognitive ToolKit (previously CNTK) is growing in popularity.

By: Arif

Arif — Mon, 07 Nov 2016 18:23:07 +0000

Anyone aware of “Microsoft Cognitive ToolKit” ?
Link – https://www.microsoft.com/en-us/research/product/cognitive-toolkit/

By: John Murphy

John Murphy — Mon, 18 Jul 2016 15:34:55 +0000

In reply to Wendy. deeplearning4j is another great framework for deep learning. The frameworks presented here are not comprehensive.

By: Wendy

Wendy — Sun, 03 Jul 2016 13:53:48 +0000

I like the deeplearning4j by Skymind, was found many corporate using it

By: John Murphy

John Murphy — Fri, 03 Jun 2016 15:03:32 +0000

In reply to Luis Goncalves.

Chainer looks rather useful. The debugging messages from Theano are not usually immediately helpful.
Keras was created in order to have a framework with the look and feel of Torch, but in Python.
For those interested, see http://docs.chainer.org/en/stable/index.html

By: Luis Goncalves

Luis Goncalves — Wed, 30 Mar 2016 17:58:20 +0000

There’s another framework which I’ve recently discovered which is remarkably much easier to use than, say, Theano.

Unlike other frameworks which compile the graph first, and then execute the generated code, Chainer is all Python, and the graph is produced at run-time .. the code you write is the code that runs (it still runs fast, sometimes faster than Theano, Torch, according to some benchmarks, on CPU and on GPU). This makes it incredibly easier to debug.

My own experience has been to try to learn Theano for two weeks and still struggling with it, in stark contrast to Chainer, which I discovered afterwards, and which I could understand quite deeply and use confidently after just *one afternoon* of reading their documentation!!! It’s beautiful, and I hope it catches on!!