Ronan Collobert (collober [at]

What's Torch ?

It's a machine-learning library, written in simple C++ and distributed now under a BSD license.
Torch is currently developed at Idiap Research Institute, in Switzerland mountains.

Update: You might want to have a look at Torch5. We will update this website soon...

The Torch Team
Ronan Samy Johnny
Ronan Collobert
(The Guru)
Samy Bengio
(Distributions Wizard)
Johnny Mariéthoz
(Capt'ain Debug)

Release 3.1
August 11, 2004

Torch 3 Vision

A full additional package for machine learning applied to vision applications is now available.
Have a look here.


If you are using Torch for your scientific research, please refer to the following paper:
R. Collobert, S. Bengio, and J. Mariéthoz. Torch: a modular machine learning software library. Technical Report IDIAP-RR 02-46, IDIAP, 2002.


  • A lot of things in gradient machines, that is, machines which could be learned with a gradient descent. This includes multi-layered perceptrons, radial basis functions, mixtures of experts, convolutional networks and even time-delay neural networks. In fact a lot of "modules" are available that you can plug as you want to get what you need.
  • Support vector machines, in classification and regression. As fast as the old stand-alone program SVMTorch II, but with the powerful environment of the library.
  • Ensemble models such as bagging or adaboost.
  • Non-parametric models such as K-nearest-neighbors, Parzen regression and Parzen density estimator.
  • Distributions stuff, like Kmeans, Gaussian mixture models, hidden Markov models, input-output hidden Markov models, and Bayes classifier.
  • Speech recognition tools (Embedded training and large vocabulary decoding).


Torch3 has been successfully tested on Linux, SunOS, FreeBSD, OSF1, Mac OS X and even MS Windows.


Seven reasons to use Torch3:
  • It's free
  • It's modular
  • It's growing
  • It has been though to be efficient
  • It has been designed to be intuitive
  • It contains a lot of state-of-the-art algorithms
  • Even if you don't like coding, it contains nice examples
You cannot imagine what we are doing for you!


  • Torch3 is free, distributed under a BSD license. That's pretty cool for you, but we need your feedback. Cite Torch in your papers! Advertise it! Send us a mail, an e-mail, something which says that you're using the library, and may encourage us. A nice postcard, a picture from your country (or of your girlfriends!), or even money. We are reachable at the following address:
    Ronan Collobert and Samy Bengio and Johnny Mariéthoz
    Idiap Research Institute
    Centre du Parc
    Rue Marconi 19
    CP 592
    1920 Martigny
  • If you have implemented a new machine algorithm or something wich could be cool to be added to the library, send us your code! We need your contribution!