With v3.0, DIGITS now supports Torch7 as an optional alternative backend to Caffe.
NOTE: Torch support is still experimental!
If you don't need a new version or custom build of Torch, you can still use deb packages to install the latest release. Follow these instructions to gain access to the required repositories, and then use this command to install:
% sudo apt-get install torch7-nvOtherwise, follow these instructions to build from source.
- Prerequisites
- Torch installer
- Luarocks dependencies
- LMDB support
- Getting Started With Torch7 in DIGITS
To install the CUDA toolkit, first get access to the required repositories by following these instructions. Then install the toolkit with this command:
% sudo apt-get install cuda-toolkit-7-5Any CUDA toolkit >= 6.5 should work.
You can also install cuDNN via deb packages:
% sudo apt-get install libcudnn4-devFollow these instructions to install Torch7 on Mac OS X and Ubuntu 12+:
http://torch.ch/docs/getting-started.html
After installing Torch, you may consider refreshing your bash by doing source ~/.bashrc. This will conveniently add the Torch th executable to your executable search path, which will allow DIGITS to find it automatically.
To use Torch7 in DIGITS, you need to install a few extra dependencies.
If you haven't done so already, install the HDF5 package:
% sudo apt-get install libhdf5-serial-devInstall extra Lua packages:
% luarocks install image
% luarocks install "https://raw.github.com/deepmind/torch-hdf5/master/hdf5-0-0.rockspec"NOTE: If you skip this section you will not be able to train Torch7 models on LMDB datasets!
Follow these instructions if you wish to use Torch7 to train networks using LMDB-encoded datasets in DIGITS. You may skip this section if you wish to only use HDF5-encoded datasets: LMDB installation instructions
Follow these instructions for information on getting started with Torch7 in DIGITS.