deeplearning2015:torch_guide

* Torch and Friends *

You can find Torch's Github here https://github.com/torch/torch7

You can find an install guide here http://torch.ch/docs/getting-started.html#_

If you're installing Torch on your own local machine following those instructions should probably be enough. I still recommended reading through the install scripts, so that you know what it's going to install.

Torch's main thing is luarocks, which lets you very quickly install packages.

When you look at the install script here https://github.com/torch/distro/blob/master/install.sh you can see, that it basically compiles Torch, installs it and then uses luarocks to install a lot of packages. Again I'd recommend reading through this script.

When installing this on CIMS or some other remote machine on which you're not root, you will probably want to install this locally by making use of the PREFIX and getting that to be set to your home folder. You might want to also load modules “cuda” or “cmake”. As with anything, you can check it's location with “which” so you know, which binary your PATH currently prefers. Please also look through the CIMS wiki here. As always, ask on Piazza, if you are stuck.

The install script looks for nvcc and nvidia-smi, if it will install cunn, cutorch and all these GPU enabled libraries. Thus make sure you have CUDA installed, if you have an NVIDIA gpu. The toolkit here includes the compiler https://developer.nvidia.com/cuda-toolkit .

Now there is also iTorch, which is an iPython kernel and thus allows you to use an iPython REPL and iPython notebook, but with Torch. You can find it here https://github.com/facebook/iTorch . I'd recommend this as an IDE for developing Torch locally, but feel free to use other tools as well. Make sure you install all the prerequisites. For making use of a remote (through a network) iTorch notebook install, make sure you add to the torch profile the corresponding file (ipython_notebook_config.py) as described here http://ipython.org/ipython-doc/1/interactive/public_server.html .

Aside from that, you can find many little tools such as

.mat to torch file conversion
https://github.com/clementfarabet/lua---mattorch
another torch ide
https://github.com/soumith/zbs-torch
debugger
https://github.com/pkulchenko/MobDebug collection of lua utilities from Facebook, specifically notice their debugger
https://github.com/facebook/fblualib

Tutorials
http://code.madbits.com/wiki/doku.php?id=tutorial

References
http://torch.ch/docs/package-docs.html#_
http://torch.ch/docs/five-simple-examples.html#_
https://github.com/torch/torch7/wiki/Cheatsheet

Compiling packages not available via luarocks
Usually these packages come with install instructions, but you can also enter the package and use “luarocks make”.

As always, if you would like some software installed on Mercer, mention it on Piazza. But always try a local install first.