Torch7

Torch7 provides a Matlab-like environment for state-of-the-art machine learning algorithms. It is easy to use and provides a very efficient implementation, thanks to an easy and fast scripting language (Lua) and an underlying C/OpenMP/CUDA implementation.


Torch7's plotting capabilities.
related read
Torch7: A Matlab-like Environment for Machines Learning

Torch7 is not officially released yet, but is in a close to final stage now. Torch7 supersedes Torch5, and for those who were using my XLearn packages, they have been ported to Torch7, and are now distributed as Torch packages (installable via Torch's own package manager, torch-pkg).

If you want to use Torch7, and cannot wait for the official release, you can easily get started by following the installation instructions on this temporary page. In a couple of lines, it amounts to: 

$ git clone https://github.com/andresy/torch
$ cd torch
$ mkdir build; cd build
$ cmake ..
$ make
$ [sudo] make install
In terms of code base, Torch7, and all our 3rd-party modules are hosted on Gihub. Here's a non exhaustive list: You can find all the others by going to my Github page. Let me know if you're interested in contributing to any of them!

Note: all of the packages listed above are packaged as Torch packages. Given a proper Torch7 install, you can easily install any of them, like this: 

$ torch-pkg install image imgraph parallel

Torch7 is the official successor of the very cool Torch5, an original work from Ronan Collobert. It is now developed and maintained by Ronan Collobert, Koray Kavukcuoglu and I. Here is a short paper that describes Torch7. It will be presented at the Big Learn workshop this year. One the coolest features of Torch7 is speed:


Torch7 is fast. Benchmarks of Torch7 (red stripes) versus Theano (solid green), while training various neural networks architectures with SGD. We considered a single CPU core, OpenMP with 4 cores and GPU alternatives. Performance is given in number of examples processed by second (higher is better). "batch" means 60 examples were fed at the time when training with SGD.