best cuda machine learning library

which machine learning library has the best support for using CUDA natively? Can you please answer on the following parameters:

price or ease of avilability
native CUDA support
good coverage of machine learning algorithms (esp deep learning)

I know matlab has parallel computing toolbox, but is too costly for student or amatuer use.

asked Jul 02 '13 at 06:19

Arun Kumar
286●10●10●16

3 Answers:

oldest newest most voted

I think you will like Theano.
1. It's free and open source.
2. It has CPU and GPU (CUDA) optimization.
3. They provide tutorials for several learning algorithms on their website.

Also, Theano's community is very helpful. For example, Frédéric Bastien, one of Theano's original developers, answered a couple of questions of mine on their user's group.
You can also find several references for Theano here on MetaOptimize...

answered Jul 02 '13 at 07:45

edersantana
155●2●5●9

edited Jul 02 '13 at 07:46

There's also pylearn2, http://deeplearning.net/software/pylearn2/ with the vision of making the use of Theano and other frameworks easier and faster.

(Jul 02 '13 at 07:50) edersantana

Thanks ! Theano looks like it! What set up have you used?

(Jul 03 '13 at 00:24) Arun Kumar

Are you acquitted with python? You can use "pip install Theano" on command line to install it. Other details can be found here http://deeplearning.net/software/theano/install.html . I heard that people got some issues while using Theano+GPU on the retina display. But they overcame it with the Enthought python distribution. Remember, you can always refer to the user group if you have problems. If you got any other question for MetaOptimize, create it's own thread.

(Jul 05 '13 at 12:59) edersantana

I prefer Theano and Torch . If you are familiar with Matlab, you might find Torch easy to use since the programming language used in Torch is Lua - a Matlab-like language. Both Theano and Torch cover deep learning algorithms and they are open-sourced.

answered Jul 02 '13 at 07:47