Skip to content

Latest commit

 

History

History
116 lines (87 loc) · 4.24 KB

UbuntuInstall.md

File metadata and controls

116 lines (87 loc) · 4.24 KB
Raw

Ubuntu Installation

Deb packages for major releases (i.e. v3.0 and v4.0 but not v3.1) are provided for easy installation on Ubuntu 14.04 and 16.04. If these packages don't meet your needs, then you can follow these instructions to build DIGITS and its dependencies from source.

Prerequisites

You need an NVIDIA driver (details and instructions).

Run the following commands to get access to some package repositories:

# For Ubuntu 14.04
CUDA_REPO_PKG=http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1404/x86_64/cuda-repo-ubuntu1404_8.0.61-1_amd64.deb
ML_REPO_PKG=http://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1404/x86_64/nvidia-machine-learning-repo-ubuntu1404_4.0-2_amd64.deb

# For Ubuntu 16.04
CUDA_REPO_PKG=http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64/cuda-repo-ubuntu1604_8.0.61-1_amd64.deb
ML_REPO_PKG=http://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1604/x86_64/nvidia-machine-learning-repo-ubuntu1604_1.0.0-1_amd64.deb

# Install repo packages
wget "$CUDA_REPO_PKG" -O /tmp/cuda-repo.deb && sudo dpkg -i /tmp/cuda-repo.deb && rm -f /tmp/cuda-repo.deb
wget "$ML_REPO_PKG" -O /tmp/ml-repo.deb && sudo dpkg -i /tmp/ml-repo.deb && rm -f /tmp/ml-repo.deb

# Download new list of packages
sudo apt-get update

Installation

Now that you have access to all the packages you need, installation is simple.

sudo apt-get install digits

Through package dependency chains, this installs digits, caffe-nv, torch7-nv, libcudnn and many more packages for you automatically.

Getting started

The DIGITS server should now be running at http://localhost/. See comments below if you run into any issues.

Now that you're up and running, check out the Getting Started Guide.

Troubleshooting

Configuration

If you have another server running on port 80 already, you may need to reconfigure DIGITS to use a different port.

sudo dpkg-reconfigure digits

All other configuration is done with environment variables. See Configuration.md for detailed information about which variables you can change.

  • Ubuntu 14.04

    • Edit /etc/init/digits.conf
    • Add/remove/edit lines that start with env
    • Restart with sudo service digits restart

  • Ubuntu 16.04

    • Edit /lib/systemd/system/digits.service
    • Add/remove/edit lines that start with Environment= in the [Service] section
    • Restart with sudo systemctl daemon-reload && sudo systemctl restart digits

Driver installations

If you try to install a new driver while the DIGITS server is running, you'll get an error about CUDA being in use. Shut down the server before installing a driver, and then restart it afterwards.

Ubuntu 14.04:

sudo service digits stop
# (install driver)
sudo service digits start

Ubuntu 16.04:

sudo systemctl stop digits
# (install driver)
sudo systemctl start digits

Permissions

The DIGITS server runs as www-data, so keep in mind that prebuilt LMDB datasets used for generic models need to be readable by the www-data user. In particular, the entire chain of directories from / to your data must be readable by www-data.

Torch and cusparse

There is at least one Torch package which is missing a required dependency on cusparse. If you see this error:

/usr/share/lua/5.1/cunn/THCUNN.lua:7: libcusparse.so.7.5: cannot open shared object file: No such file or directory

The simplest fix is to manually install the missing library:

sudo apt-get install cuda-cusparse-7-5
sudo ldconfig

Torch and HDF5

There is at least one Torch package which is missing a required dependency on libhdf5-dev. If you see this error:

ERROR: /usr/share/lua/5.1/trepl/init.lua:384: /usr/share/lua/5.1/trepl/init.lua:384: /usr/share/lua/5.1/hdf5/ffi.lua:29: libhdf5.so: cannot open shared object file: No such file or directory

The simplest fix is to manually install the missing library:

sudo apt-get install libhdf5-dev
sudo ldconfig

Other

If you run into an issue not addressed here, try searching through the GitHub issues and/or the user group.