Setting Up CUDA + cuDNN for Theano in Ubuntu

This is my personal notes on setting up Theano with my laptop GPU. It is basically an amalgam of various sources that I pieced together to make everything work, which I will link at the end of this post.

As of this writing, this is my setup:

  • Linux Mint 18 (Cinnamon) 64-bit (based on Ubuntu 16.06)
  • NVIDIA GT740M (Kepler architecture)
  • Theano 0.8.2

NVIDIA Graphic Drivers

Linux Mint gives you an option to install the drivers from the settings, but it may be dated. To get the latest drivers, you may install the drivers via PPA:

IMPORTANT: You need to install the drivers first, before installing CUDA. The order is very important, since CUDA checks what version of the graphic driver you are using and installs accordingly. On a related note – should you upgrade/downgrade your graphics driver, you will need to install CUDA again. I emphasize this, because should you fail to do so, the errors that proceed it gives you no indication whatsoever that you screwed this step up.


Choose CUDA for download in As of this writing the latest version is CUDA 7.5, so that is what I downloaded. After you download, there are instructions that they suggest you run. Not all of them will work. So follow these suggested steps instead:

Open a terminal in the download directory and enter the first command they suggested for you in the downloads site. It should look like this:

sudo dpkg -i cuda-repo-ubuntu1504-7-5-local_7.5-18_amd64.deb

Change your /var/cuda-repo-7-5-local/Release to the following:

Origin: NVIDIA
Architecture: repogenstagetemp
 51483bc34577facd49f0fbc8c396aea0 75379 Packages
 4ef963dfa4276be01db8e7bf7d8a4f12 21448 Packages.gz
 532b1bb3b392b9083de4445dab2639b36865d7df1f610aeef8961a3c6f304d8a 75379 Packages
 2e48cc13b6cc5856c9c6f628c6fe8088ef62ed664e9e0046fc72819269f7432c 21448 Packages.gz

Run (ignoring warnings about invalid signatures, and you’re done):

sudo apt-get update

Then run:

sudo apt-get install cuda

Keep an eye out the output. There should not be any errors. This will install CUDA in /usr/local/cuda/

Add CUDA To Environment Paths

Open ~/.bashrc  and append the following (you may need to do this in sudo mode):

export PATH=/usr/local/cuda/bin:$PATH 
export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH

Once you save, enter in a terminal:

sudo source ~/.bashrc

This will add the CUDA executables to the environment paths. Note that currently opened terminals will not have CUDA added to the environment paths. You will need to restart them (open and close) for changes to take affect.

And then you can open a new terminal and type nvcc (Nvidia CUDA Compiler) to see whether the environment is set correctly. It should not output any errors.

Solving gcc/g++ Incompatibilities

CUDA requires a compatible C/C++ compiler to work. The one that comes bundled with Ubuntu isn’t. To fix this, enter the following:

sudo apt-get install gcc-4.9 g++-4.9

Then we may establish a soft link of the specific version for the CUDA binaries folder:

sudo ln -s /usr/bin/gcc-4.9 /usr/local/cuda/bin/gcc
sudo ln -s /usr/bin/g++-4.9 /usr/local/cuda/bin/g++

IMPORTANT! Now, if you run import theano for the first time with the THEANO_FLAGS environment variable containing device=gpu, theano complains that CUDA is not available. To run any python script that uses Theano, you need to prepend the command with THEANO_FLAGS=device=gpu,nvcc.flags=-D_FORCE_INLINES. All python scripts executed here will be using this workaround. Alternatively there is a fix here: (thanks Anonoz for the suggestion).


Now running the following line:

THEANO_FLAGS=device=gpu,nvcc.flags=-D_FORCE_INLINES python -c "import theano; print(theano.sandbox.cuda.device_properties(0))"

Should give you something like this:

Using gpu device 0: GeForce GT 740M (CNMeM is disabled, CuDNN not available)
{'major': 3, 'tccDriver': 0, 'kernelExecTimeoutEnabled': 1, 'deviceOverlap': 1, 'driverVersion': 8000, 'warpSize': 32, 'concurrentKernels': 1, 'maxThreadsPerBlock': 1024, 'computeMode': 0, 'canMapHostMemory': 1, 'maxGridSize2': 65535, 'maxGridSize1': 65535, 'maxGridSize0': 2147483647, 'integrated': 0, 'minor': 0, 'ECCEnabled': 0, 'runtimeVersion': 7050, 'textureAlignment': 512, 'multiProcessorCount': 2, 'clockRate': 895000, 'totalConstMem': 65536, 'name': 'GeForce GT 740M', 'memPitch': 2147483647, 'maxThreadsDim1': 1024, 'maxThreadsDim0': 1024, 'maxThreadsDim2': 64, 'coresCount': -2, 'sharedMemPerBlock': 49152, 'regsPerBlock': 65536}


NVIDIA provides a library for common neural network operations that especially speeds up Convolutional Neural Networks (CNNs). For Lasagne, it is necessary that you install this to get a convnet to work. It can be obtained from NVIDIA (after registering as a developer):

Don’t expect an instant email upon registration. For some reason it takes quite a while for them to send that email. I waited about 30 minutes.

Once you are in, choose version 4. That’s the one currently supported by Theano.

To install it, copy the *.h files to /usr/local/cuda/include and the lib* files to /usr/local/cuda/lib64

To check whether it is installed, run

THEANO_FLAGS=device=gpu,nvcc.flags=-D_FORCE_INLINES python -c "from theano.sandbox.cuda.dnn import dnn_available as d; print(d() or d.msg)"

It will print True if everything is fine, or an error message otherwise. There are no additional steps required for Theano to make use of cuDNN.

Again, if everything if successful, you run your python scripts as such (the following is deep_q_rl, a Theano-based implementation of Deep Q-learning using Lasagne):

 THEANO_FLAGS=device=gpu,nvcc.flags=-D_FORCE_INLINES python --rom breakout

References (in order):

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

This site uses Akismet to reduce spam. Learn how your comment data is processed.