Software in Quadcorn
posted on August 17, 2016

We generally followed the instructions in Setting up a Deep Learning Machine from Scratch.

Installing NVidia Drivers

Before installing the drivers, make sure there are no nvidia drivers already installed by doing the following:

$ sudo apt-get --purge remove nvidia-*

And reboot. Then download NVidia drivers from here and install them according to instructions.

Installing CUDA

First, install some AMD drivers that will be used by CUDA.

$ sudo apt-get install fglrx-updates

After installation, reboot the computer.

Next, stop the SSH X11 forwarding by modifying /etc/ssh/sshd_config:

$ sudo nano /etc/ssh/sshd_config

Find the following line:

X11Forwarding yes

and change it to:

X11Forwarding no

After saving the changes, restart the ssh service by typing:

$ sudo service ssh restart

Now, stop the default X server by typing:

$ sudo service mdm stop

Now we can install CUDA. The latest supported version is 7.5 which can be downloaded from here. Select Linux > X86_64 > Ubuntu > 14.04 > runfile (local). Follow the instructions on the page to complete the installation.

After finishing the installation of CUDA, undo the change that you made to /etc/ssh/sshd_config, and restart the ssh server using sudo service ssh restart. Also, start the X server again: sudo service mdm start

Installing cuDNN

Download cuDNN from here. Select a version which is compatible with CUDA. After downloading, do the following:

$ cd cuda
$ sudo cp */*.h /usr/local/cuda/include/
$ sudo cp */libcudnn* /usr/local/cuda/lib64/
$ sudo chmod a+r /usr/local/cuda/lib64/libcudnn*


To install for all users

$ sudo -i

then install package. git clone in ~\repos if need.

To install only for user then add --user at the end. E.g.:

$ python install --user

Here are some specific instructions:

$ pip install theano

Edit/create ~/.theanorc:

floatX = float32
device = gpuX

choose as gpuX your assigned gpu, there will be random inspections. Grad students all share gpu1. klab uses gpu0

$ export TF_BINARY_URL=
$ pip install --ignore-installed --upgrade $TF_BINARY_URL

Note: right now tensorflow is not working because we are using cuDNN v5. Either wait for it to be compatible or downgrade cuDNN to v4 when we really need to use tensorflow (don’t do it). Or try to compile/install tensorflow from source (good luck with that).

$ pip install keras

See here to choose tensorflow and theano in Keras

$ git clone --recursive
$ cd xgboost; make -j4
$ cd python-package
$ python install
$ conda install opencv