These instruction was done for Ubuntu 11.04, 11.10 and 12.04. You can probably do something similar on older computer.
It is possible to have a faster installation of Theano than the one these instructions will provide, but this will make the installation more complicated and/or may require that you buy software. This is a simple set of installation instructions that will leave you with a relatively well-optimized version that uses only free software. With more work or by investing money (i.e. buying a license to a proprietary BLAS implementation), it is possible to gain further performance.
If you are behind a proxy, you must do some extra configuration steps before starting the installation. You must set the environment variable http_proxy to the proxy address. Using bash this is accomplished with the command export http_proxy="http://user:email@example.com:port/" You can also provide the --proxy=[user:pass@]url:port parameter to pip. The [user:pass@] portion is optional.
We use pip for 2 reasons. First, it allows “import module; module.test()” to work correctly. Second, the installation of NumPy 1.6 or 1.6.1 with easy_install raises an ImportError at the end of the installation. To my knowledge we can ignore this error, but this is not completely safe. easy_install with NumPy 1.5.1 does not raise this error.
This page describes how to install Theano for Python 2. If you have installed Python 3 on your system, maybe you need to change the command pip to pip-2.7 to specify to install it for Python 2, as sometimes the pip command refers to the Python 3 version.
The development version of Theano supports Python 3.3 and probably supports Python 3.2, but we do not test on it.
If the packages libatlas3gf-base or libatlas-dev are already installed, there will be problems as they conflict with libopenblas-dev. If you see NumPy errors, the simplest is to remove libopenblas-dev and its dependency libopenblas-base like this: sudo apt-get remove libopenblas-base. The ideal would be that you remove libatlas3gf-base and libatlas-dev, but you will need to reinstall python-numpy, python-scipy and all other packages that used it. OpenBLAS is faster then ATLAS most of the time and it allows to control the number of threads used during the execution.
If you have error that contain “gfortran” in it, like this one:
ImportError: (‘/home/Nick/.theano/compiledir_Linux-2.6.35-31-generic-x86_64-with-Ubuntu-10.10-maverick–2.6.6/tmpIhWJaI/0c99c52c82f7ddc775109a06ca04b360.so: undefined symbol: _gfortran_st_write_done’
The problem is probably that NumPy is linked with a different blas then then one currently available (probably ATLAS). There is 2 possible fixes:
1) is better as OpenBLAS is faster then ATLAS and NumPy is probably already linked with it. So you won’t need any other change in Theano files or Theano configuration.
- NumPy (~30s): python -c "import numpy; numpy.test()"
- SciPy (~1m): python -c "import scipy; scipy.test()"
- Theano (~30m): python -c "import theano; theano.test()"
NumPy 1.6.2, 1.7.0 and 1.7.1, have a bug where it marks some ndarrays as not aligned. Theano does not support unaligned arrays, and raises an Exception when that happens. This can cause one test to fail with an unaligned error with those versions of NumPy. You can ignore that test error as at worst, your code will crash. If this happens, you can install another NumPy version to fix this problem. NumPy 1.6.2 is used in Ubuntu 12.10 and NumPy 1.7.1 is used in Ubuntu 13.04.
It is recommended to test your Theano/BLAS integration. There are many versions of BLAS that exist and there can be up to 10x speed difference between them. Also, having Theano link directly against BLAS instead of using NumPy/SciPy as an intermediate layer reduces the computational overhead. This is important for BLAS calls to ger, gemv and small gemm operations (automatically called when needed when you use dot()). To run the Theano/BLAS speed test:
python `python -c "import os, theano; print os.path.dirname(theano.__file__)"`/misc/check_blas.py
This will print a table with different versions of BLAS/numbers of threads on multiple CPUs and GPUs. It will also print some Theano/NumPy configuration information. Then, it will print the running time of the same benchmarks for your installation. Try to find a CPU similar to yours in the table, and check that the single-threaded timings are roughly the same.
Theano should link to a parallel version of Blas and use all cores when possible. By default it should use all cores. Set the environment variable “OMP_NUM_THREADS=N” to specify to use N threads.
If you followed these installation instructions, you can execute this command to update only Theano:
sudo pip install --upgrade --no-deps theano
If you want to also installed NumPy/SciPy with pip instead of the system package, you can run this:
sudo pip install --upgrade theano
Do like in the section “Updating Theano”, but use git+git://github.com/Theano/Theano.git instead of theano.
The openblas included in Ubuntu is limited to 2 threads. If you want to use more cores at the same time, you will need to compile it yourself. Here is some code that will help you.
# remove openblas if you installed it sudo apt-get remove libopenblas-base # Download the development version of OpenBLAS git clone git://github.com/xianyi/OpenBLAS cd OpenBLAS make FC=gfortran sudo make PREFIX=/usr/local/ install cd /usr/local/lib ln -s libopenblas.so /usr/lib/libblas.so ln -s libopenblas.so.0 /usr/lib/libblas.so.3gf
Basic configuration for the GPU Using the GPU.
Ubuntu 11.10/12.04 (probably work on 11.04 too):
sudo apt-add-repository ppa:ubuntu-x-swat/x-updates sudo apt-get update sudo apt-get install nvidia-current
Then you need to fetch latest CUDA tool kit (download ubuntu 11.04 32/64bit package) from here.
For the run installed (the only one available for CUDA 5.0 and older), you install it like this:
chmod a+x XXX.sh sudo ./XXX.sh
Since CUDA 5.5, Nvidia provide a DEB package. If you don’t know how to intall it, just double click on it from the graphical interface. It should ask if you want to install it.
You must reboot the computer after the driver installation. To test that it was loaded correctly after the reboot, run the command nvidia-smi from the command line.
You probably need to change the default version of gcc as explained by Benjamin J. McCann:
sudo apt-get install nvidia-cuda-toolkit g++-4.4 gcc-4.4 # On Ubuntu 11.10 and 12.04, you probably need to change gcc-4.5 to gcc-4.6 on the next line. sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-4.5 40 --slave /usr/bin/g++ g++ /usr/bin/g++-4.5 sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-4.4 60 --slave /usr/bin/g++ g++ /usr/bin/g++-4.4 sudo update-alternatives --config gcc
THEANO_FLAGS=floatX=float32,device=gpu python /usr/lib/python2.*/site-packages/theano/misc/check_blas.py
Ubuntu 10.04 LTS: default gcc version 4.4.3. gcc 4.1.2, 4.3.4 availables.
Ubuntu 11.04: default gcc version 4.5.2. gcc 4.4.5 availables.
Ubuntu 11.10: default gcc version 4.6.1. gcc 4.4.6 and 4.5.3 availables.
Ubuntu 12.04 LTS: default gcc version 4.6.3. gcc 4.4.7 and 4.5.3 availables.
Ubuntu 12.10: default gcc version 4.7.2. gcc 4.4.7, 4.5.4 and 4.6.3 availables.