Installation of Theano on Windows

These instructions show step-by-step how to install Theano and required dependencies on a 32- or 64-bit system using freely available tools and compilers.

Installing Dependencies

Note

Command lines listed below are assumed to be run in a Windows prompt: To open the prompt on Windows <= 7, click Start and type the cmd command to launch a command window. In Windows 8, go to the Start screen and type command or cmd.

Theano currently works on Windows, however it requires compilers for C/C++ (for Python 2.7 family this has to be Microsoft Visual Studio 2008 compiler), CUDA (CUDA v5.5 is required as it is the latest version supporting Visual Studio 2008), and GCC (for non-CUDA C code generated by Theano).

Visual Studio and CUDA

Unfortunately Microsoft recently stopped distributing Visual Studio Express 2008 (the compilers required for Python 2.7 are provided, though), therefore we require a temporary install of Visual Studio Express 2010 to be able to install CUDA (its installer requires a Visual Studio installation to proceed). Afterwards, the Visual Studio 2010 can be safely removed. If someone knows how to install CUDA 5.5 without a proper Visual Studio installation, please let us know.

First we need to install Microsoft Visual Studio 2010 Express, which is required to install CUDA. You can download it from Visual Studio Express. Please install the Visual C version. We have downloaded the all-in-one CD, extracted it using 7zip, and ran the installer at VCExpress\setup.exe.

If you want a 64bit Python installation, Visual Studio 2010 Express doesn’t provide a 64bit compiler. To get one download and install the Windows Software Development Kit version 7.1.

Now you have a running (and free even for commercial use) installation of MSVS2010 IDE with 32- and 64-bit compilers.

Once Visual Studio is installed, you can install CUDA. We recommend CUDA 5.5, as it supports MSVC 2008. Download the CUDA installer from the CUDA archive. Be sure to get 32-, or 64-bit version depending on your system configuration.

Once CUDA is installed you can remove VisualStudio Express 2010.

Finally, grab the Microsoft Visual C++ Compiler for Python 2.7. It provides the now-obsolete compilers form Visual Studio 2008 that are required for compatibility with Python 2.7. To install the package for all users please:

  1. open an administrator’s console (got to start, then type cmd, right-click on the command prompt icon and select run as administrator)
  2. cd to your downloads directory and execute msiexec /i VCForPython27.msi ALLUSERS=1

The package will be installed to C:\Program Files (x86)\Common Files\Microsoft\Visual C++ for Python\9.0.

Finally download the stdint.h header from here and save it as C:\Program Files (x86)\Common Files\Microsoft\Visual C++ for Python\9.0\VC\include\stdint.h.

GCC

Theano C code compiler currently requires a GCC installation. We have used the build TDM GCC which is provided for both 32- and 64-bit platforms. A few caveats to watch for during installation:

  1. Install to a directory without spaces (we have placed it in C:\SciSoft\TDM-GCC-64)
  2. If you don’t want to clutter your system PATH un-check add to path option.
  3. Enable OpenMP support by checking the option openmp support option.

Scientific Python distribution

We highly recommend the Pierre Raybaut’s WinPython distribution - it is compiled for both 32- and 64-bit systems, links against the fast MKL BLAS implementation, supports installation of auxiliary packages from Chris Gohlke and is free.

WinPython also allows for a portable installation and doesn’t clutter your main system PATH. We have installed it to c:\SciSoft\WinPython-64bit-2.7.9.4.

Alternative in academia: EPD

If you are working in academia, an easy way to install most of the dependencies is to install Enthought Python Distribution (EPD). If you are affiliated with a university (as student or employee), you can download the installation for free. EPD installation includes, in particular, Python (and the development headers), NumPy, SciPy, nose, sphinx, easy_install, pydot (but not Graphviz, which is necessary for it to work), g++, and the MKL implementation of blas.

If you want to use the iPython shell, you should first try to import NumPy in it:

C:\Users\user>ipython
[...]
In [1]: import numpy

If you see an error message telling you that DLL load failed, that is probably due to a bug in the script launching ipython. If C:\SciSoft\Python27 is the directory where you installed EPD, edit C:\SciSoft\Python27\Scripts\ipython.bat, there should be a line saying:

set path="C:\SciSoft\Python27";%path%

Remove the quotes around Python27, leading to:

set path=C:\SciSoft\Python27;%path%

Then, it should work in all new terminals.

pip is not included in EPD, but you can simply install it with:

easy_install pip

Alternative: Canopy

Canopy is another software that installs all Theano dependencies. If you are affiliated with a university (as student or employee), you can download the installation for free.

  • Install Canopy x64, and update it to the latest version (Help / Software updates...), as older Canopy versions have trouble installing pip.
  • Then install pip from Canopy Package Manager.
  • In the Windows Prompt, type pip install theano.
  • In Canopy Package Manager, search and install packages “mingw 4.5.2” and “libpython 1.2”
  • (Needed only for Theano 0.6rc3 or earlier) The “libpython 1.2” package installs files libpython27.a and libmsvcr90.a to C:\Users\<USER>\AppData\Local\Enthought\Canopy\User\libs. Copy the two files to C:\Users\<USER>\AppData\Local\Enthought\Canopy\App\appdata\canopy-1.0.0.1160.win-x86_64libs.
  • (Needed only for Theano 0.6rc3 or earlier) Set the Theano flags blas.ldflags=-LC:\Users\<USER>\AppData\Local\Enthought\Canopy\App\appdata\canopy-1.0.0.1160.win-x86_64\Scripts -lmk2_core -lmk2_intel_thread -lmk2_rt.

Alternative: Anaconda

ContinuumIO is providing a free Python distribution for Windows (32-bit and 64-bit), including all dependencies of Theano. If you are not eligible for a download of EPD or Canopy (via a commercial, or free academic licence), this is the easiest way to install Theano’s dependencies. Simply download and execute the installer from Anaconda download page, and execute the following in Windows command line:

$ conda install mingw libpython

Alternative: Python(x,y)

If you do not have a commercial licence of EPD, and are not eligible to a free academic licence, and neither Python nor MinGW is installed on your computer, you can install most dependencies of Theano with Python(x,y). It is a single installation file that contains additional packages like NumPy, SciPy, IPython, Matplotlib, MinGW, Nose, etc. Note however that there is no 64 bit version currently available. You can keep the default install options, except that the installation directory should not contain any blank space (in particular, do not install it into C:\Program Files).

Alternative: manual installation

The following instructions provide steps for manual installation of all Theano dependencies. Note that it should be possible to run Theano with Cygwin instead of MinGW, but this has not been tested yet.

  • For 32 bit MinGW: from the MinGW files, download the latest version of the Automated MinGW Installer (mingw-get-inst) and install it (you should install all optional components, except the Objective C and Ada compilers which are not needed).

  • For 64 bit MinGW (note that manual installation for 64 bit is experimental): download the latest version of MinGW-w64 from the project’s releases page, and extract it for instance to C:\mingw64. Also download MSYS from this page (although it is a 32-bit version of MSYS, this does not matter since it is only a convenience tool). Extract MSYS into the same folder, so that, for instance, you end up with C:\mingw64\msys. Run C:\mingw64\msys\msys.bat and in the MSYS shell, type

    sh /postinstall/pi.sh
    

    and answer the few questions so that MSYS is properly linked to your MinGW install.

  • It is recommended to set your MSYS home to be the same as your Windows home directory. This will avoid inconsistent behavior between running Theano in a Windows command prompt vs. a MSYS shell. One way to do this without setting a global Windows HOME environment variable (which may affect other programs) is to edit your msys.bat file (found e.g. under C:\MinGW\msys\1.0 or C:\mingw64\msys) and add the following line at the beginning (note that you may need to use e.g. Wordpad to edit this file, since Notepad gets confused by Unix-style line breaks):

    set HOME=%USERPROFILE%
    
  • If you do not have them already, install the latest versions of Python 2.x and corresponding NumPy then SciPy packages (simply use the executable installers). Note that there are currently no official 64 bit releases of NumPy and SciPy, but you can find unofficial builds here.

  • Ensure that the Python installation directory and its Scripts sub-directory are in your system path. This may be done by modifying the global PATH Windows environment variables, or by creating a .profile file in your MinGW home, containing a line like export PATH=$PATH:/c/Python27:/c/Python27/Scripts (note that the latter will work only when you run Theano from an MSYS shell).

  • If you are installing the 64 bit version, you will need the following hack to be able to compile Theano files with GCC (skip this step if you are using the 32 bit version). In a temporary work directory, copy python27.dll (found in C:\\Windows\\System32) as well as python27.def. Edit python27.def and replace Py_InitModule4 with Py_InitModule4_64. Then open an MSYS shell, go to this temporary directory, and run:

    dlltool --dllname python27.dll --def python27.def --output-lib libpython27.a
    

    Finally, copy the libpython27.a file that was generated into your C:\\Python27\\libs folder.

  • In order to run Theano’s test-suite, you will need nose. After unpacking its source code (you may use 7-zip), you can build and install it from within its code directory by running the following command (either from a Windows command prompt or an MSYS shell):

    python setup.py install
    

Configuring the Environment

At this point, you should have installed all Theano dependencies. By default neither Python, GCC, nor Visual Studio was added to the PATH. Save the following shell script as c:\scisoft\env.bat to configure the system path:

REM configuration of paths
set VSFORPYTHON="C:\Program Files (x86)\Common Files\Microsoft\Visual C++ for Python\9.0"
set SCISOFT=%~dp0

REM add tdm gcc stuff
set PATH=%SCISOFT%\TDM-GCC-64\bin;%SCISOFT%\TDM-GCC-64\x86_64-w64-mingw32\bin;%PATH%

REM add winpython stuff
CALL %SCISOFT%\WinPython-64bit-2.7.9.4\scripts\env.bat

REM configure path for msvc compilers
REM for a 32 bit installation change this line to
REM CALL %VSFORPYTHON%\vcvarsall.bat
CALL %VSFORPYTHON%\vcvarsall.bat amd64

REM return a shell
cmd.exe /k

The script assumes that you installed WinPython distribution, update the winpython line otherwise. For a 32 bit installation please change the indicated line to load 32-bit Microsoft Compilers.

You can access the Python shell by double-clicking on c:\scisoft\env.bat. Please do so, and verify that the following programs are found:

  1. where gcc
  2. where gendef
  3. where cl
  4. where nvcc

Finally we need to create a link library for GCC. Open up the Python shell and cd to c:\SciSoft. Then execute:

gendef WinPython-64bit-2.7.9.4\python-2.7.9.amd64\python27.dll

dlltool --dllname python27.dll --def python27.def --output-lib WinPython-64bit-2.7.9.4\python-2.7.9.amd64\libs\libpython27.a

Installing Theano

Once the dependencies are installed, you can download and install Theano. We have found that in the long run, the Git install is the most useful, because you can update it with a single git pull command. Therefore we recommend it. However, a manual install without Git is also possible.

Git Install

Theano is hosted on GitHub, you need Git to download it. For Windows, download and install the MSYSGIT build. Open up the Git Shell in the directory in which you want to install Theano. For the bleeding-edge version execute

git clone https://github.com/Theano/Theano.git

For the latest stable release 0.7 (as of March 2015) run instead:

git clone https://github.com/Theano/Theano.git --branch rel-0.7

Either way, a folder Theano will be created with the library downloaded to it.

Manual Installation

To get the latest bleeding edge version got to Theano on GitHub and download the latest zip. Then unzip it somewhere.

Alternatively, you can check the latest release release 0.7 (as of March 2015) by going to https://github.com/Theano/Theano/releases/tag/rel-0.7 and downloading the zip.

Configuring Theano

Once you have installed Theano, open the Python Shell (e.g. c:\scisoft\env.bat if you follow the installation directories from this tutorial) and cd to it. Then run:

python setup.py develop

this step will add the Theano directory to you PYTHON_PATH environment variable.

At this stage you can check whether Theano works and is able to compile C code for CPU execution.

Create a test file containing:

import numpy as np
import time
import theano
A = np.random.rand(1000,10000).astype(theano.config.floatX)
B = np.random.rand(10000,1000).astype(theano.config.floatX)
np_start = time.time()
AB = A.dot(B)
np_end = time.time()
X,Y = theano.tensor.matrices('XY')
mf = theano.function([X,Y],X.dot(Y))
t_start = time.time()
tAB = mf(A,B)
t_end = time.time()
print "NP time: %f[s], theano time: %f[s] (times should be close when run on CPU!)" %(
                                           np_end-np_start, t_end-t_start)
print "Result difference: %f" % (np.abs(AB-tAB).max(), )
NP time: 1.480863[s], theano time: 1.475381[s] (times should be close when run on CPU!)
Result difference: 0.000000

Then run it. It should execute without problems and the Theano function should run at a speed similar to the regular NumPy multiplication. (Both Numpy and Theano should call the same BLAS routine for matrix multiplication)

Configure Theano for GPU use

Theano can be configured with a .theanorc text file (or .theanorc.txt, whichever is easier for you to create under Windows). It should be placed in the directory pointed to by the %USERPROFILE% variable. Please note, that WinPython changes it to WinPythonDir\settings (so in our system this corresponds to c:\scisoft\WinPython-64bit-2.7.9.4\settings.

To use the GPU please write the following configuration file:

[global]
device = gpu
floatX = float32

[nvcc]
flags = --use-local-env  --cl-version=2008

Rerun the simple test file and verify that it runs. Depending on you GPU, the theano function should run on the GPU much faster than the CPU matrix multiplication performed by NumPy. You can also find additional test code and useful GPU tips on the Using the GPU page.

Running Theano’s test-suite

Currently, due to memory fragmentation issue in Windows, the test-suite breaks at some point when using theano-nose, with many error messages looking like: DLL load failed: Not enough storage is available to process this command. As a workaround, you can instead run:

theano-nose --batch

This will run tests in batches of 100, which should avoid memory errors. Note that this script calls nosetests, which may require being run from within an MSYS shell if you installed Nose manually as described above.

Note

In Theano versions <= 0.5, theano-nose was not included. If you are working with such a version, you can call this command instead:

python theano/tests/run_tests_in_batch.py

Compiling a faster BLAS

If you installed Python through WinPython or EPD, Theano will automatically link with the MKL library, so you should not need to compile your own BLAS.

Note

The instructions below have not been tested in a Windows 64 bit environment.

If you want a faster and/or multi-threaded BLAS library, you can compile OpenBLAS (ATLAS may work too, but was not tested, and is usually reported to be slower and more difficult to compile – especially on Windows). OpenBLAS can be downloaded as a zip file from its website (we tested v0.2.6). To compile it, you will also need MSYS and wget (installation steps are described below).

If you already have a full install of MinGW, you should have MSYS included in it, and thus should be able to start a MinGW shell. If that is the case, you can skip the following MSYS installation steps. Note that these steps were written for Python(x,y), but should also work for other bundle Python distributions like EPD (changing paths accordingly, for instance in EPD 7.3.2 the MinGW folder is EPD7.3.2\EGG-INFO\mingw\usr\i686-w64-mingw32). To install MSYS on top of the MinGW installation included within Python(x,y), do as follows:

  • Download the mingw-get command-line installer binary.

  • Unpack its content into your pythonxy\mingw directory.

  • In a prompt (cmd), install MSYS with

    mingw-get install msys-base
    

    If mingw-get cannot be found automatically, just navigate first into the folder were it was extracted (it is found in the bin subfolder).

  • Edit pythonxy\mingw\msys\1.0\msys.bat (e.g. in Wordpad) and add as first line set HOME=%USERPROFILE%. Then create an easily accessible shortcut (e.g. on your desktop) to this file, run it and within the MSYS console, run the MSYS post-install script:

    /postinstall/pi.sh
    

    It will ask for your MinGW installation directory (e.g. c:/pythonxy/mingw; note the forward slashes).

Once you have a working MinGW/MSYS shell environment, you can go on as follows:

  1. Install wget by running the setup program you can download on the wget website. Note that this setup does not add wget into the system PATH, so you will need to modify the PATH environment variable accordingly (either in Windows or in a .profile startup file in your MinGW home). Once this is done, type wget --version in a MinGW shell to verify that it is running properly. Note also that if you are behind a proxy, you should set up your HTTP_PROXY environment variable, or use a custom wgetrc config file for wget to be able to download files.
  2. Unzip OpenBLAS and, in a MinGW shell, go into the corresponding directory.
  3. Compile OpenBLAS with:
quickbuild.win32 1>log.txt 2>err.txt

(use quickbuild.win64 for 64-bit Windows). Compilation can take a while, so now is a good time to take a break. When it is done, you should have libopenblas.dll in your OpenBLAS folder. If that is not the case, check the err.txt log for build errors.

  1. Make sure that libopenblas.dll is in a folder that is in your PATH.
  2. Modify your .theanorc (or .theanorc.txt) with ldflags = -LX:\\YYY\\ZZZ -lopenblas where X:\\YYY\\ZZZ is the path to the folder containing libopenblas.dll. This setting can also be changed in Python for testing purpose (in which case it will remain only for the duration of your Python session):
theano.config.blas.ldflags = "-LX:\\YYY\\YYY -lopenblas"
  1. To test the BLAS performance, you can run the script theano/misc/check_blas.py. Note that you may control the number of threads used by OpenBLAS with the OPENBLAS_NUM_THREADS environment variable (default behavior is to use all available cores). Here are some performance results on an Intel Core2 Duo 1.86 GHz, compared to using NumPy’s BLAS or the un-optimized standard BLAS (compiled manually from its source code). Note that we report here results for GotoBLAS2 which is the ancestor of OpenBLAS (this benchmark still needs to be updated with OpenBLAS results):

    • GotoBLAS2 (2 threads): 16s
    • NumPy (1 thread): 48s
    • Standard BLAS (un-optimized, 1 thread): 166s
    Conclusions:
    • The unoptimized standard BLAS is very slow and should not be used.
    • The Windows binaries of NumPy were compiled with ATLAS and are surprisingly fast.
    • GotoBLAS2 is even faster, in particular if you can use multiple cores.

Note

If you get a DLL load failed error message, it typically means that a required DLL was not found in the PATH. If it happens only when you are using OpenBLAS, it means it is either libopenblas.dll itself or one of its dependencies. In the case where it is a dependency, you can use the Dependency Walker utility to figure out which one.