Installing DyNet for C++¶
How to build DyNet and link it with your C++ programs.
DyNet relies on a number of external programs/libraries including CMake, Eigen, and Mercurial (to install Eigen). CMake, and Mercurial can be installed from standard repositories.
For example on Ubuntu Linux:
sudo apt-get install build-essential cmake mercurial
Or on macOS, first make sure the Apple Command Line Tools are installed, then get CMake, and Mercurial with either homebrew or macports:
xcode-select --install brew install cmake hg # Using homebrew. sudo port install cmake mercurial # Using macports.
On Windows, see Windows Support.
To compile DyNet you also need the development version of the Eigen library. If you use any of the released versions, you may get assertion failures or compile errors. If you don’t have Eigen already, you can get it easily using the following command:
hg clone https://bitbucket.org/eigen/eigen/ -r 346ecdb
The -r NUM specified a revision number that is known to work. Adventurous users can remove it and use the very latest version, at the risk of the code breaking / not compiling. On macOS, you can install the latest development of Eigen using Homebrew:
brew install --HEAD eigen
To get and build DyNet, clone the repository
git clone https://github.com/clab/dynet.git
then enter the directory and use
to generate the makefiles
cd dynet mkdir build cd build cmake .. -DEIGEN3_INCLUDE_DIR=/path/to/eigen
Then compile, where “2” can be replaced by the number of cores on your machine
make -j 2
To see that things have built properly, you can run
which will train a multilayer perceptron to predict the xor function.
If any process here fails, please see Asking Questions/Reporting Bugs for help.
GPU (CUDA) support¶
DyNet supports running programs on GPUs with CUDA. If you have CUDA
installed, you can build DyNet with GPU support by adding
-DBACKEND=cuda to your cmake options. This will result in three
libraries named “libdynet” and “libgdynet” being
created. When you want to run a program on CPU, you can link to the
“libdynet” library as shown above. When you want to run a program on
GPU, you can link to the “libgdynet” library.
(Eventually you will be able to use a single library to run on either CPU or GPU, but this is not fully implemented yet.)
When running DyNet with CUDA on GPUs, some of DyNet’s functionalities
(e.g. conv2d) will depend on the NVIDIA cuDNN libraries.
CMake will automatically detect cuDNN in the suggested installation path
by NVIDIA (i.e.
/usr/local/cuda) and enable those functionalities
If CMake is unable to find cuDNN automatically, try setting CUDNN_ROOT, such as
. However, if you don’t have cuDNN installed, those dependend functionalities will be automatically disabled and an error will be throwed during runtime if you try to use them.
Currently, DyNet supports cuDNN v5.1, future versions will also be supported soon.
DyNet can leverage Intel’s MKL library to speed up computation on the CPU. As an example, we’ve seen 3x speedup in seq2seq training when using MKL. To use MKL, include the following cmake option:
If CMake is unable to find MKL automatically, try setting MKL_ROOT, such as
One common install location is
If either MKL or MKL_ROOT are set, CMake will look for MKL.
By default, MKL will use all CPU cores. You can control how many cores MKL uses by setting the environment variable MKL_NUM_THREADS to the desired number. The following is the total time to process 250 training examples running the example encdec (on a 6 core Intel Xeon E5-1650):
encdec.exe --dynet-seed 1 --dynet-mem 1000 train-hsm.txt dev-hsm.txt
+-----------------+------------+---------+ | MKL_NUM_THREADS | Cores Used | Time(s) | +-----------------+------------+---------+ | <Without MKL> | 1 | 28.6 | | 1 | 1 | 13.3 | | 2 | 2 | 9.5 | | 3 | 3 | 8.1 | | 4 | 4 | 7.8 | | 6 | 6 | 8.2 | +-----------------+------------+---------+
As you can see, for this particular example, using MKL roughly doubles the speed of computation while still using only one core. Increasing the number of cores to 2 or 3 is quite beneficial, but beyond that there are diminishing returns or even slowdown.
Compiling with Boost¶
DyNet requires Boost for a few pieces of less-commonly-used functionality
to be enabled (unit tests and multi-processing). Boost can be enabled by using the
-DENABLE_BOOST=ON flag to
cmake. In general, DyNet will find
Boost it if it is in the standard
location. If Boost is in a non-standard location, say
you can specify the location by adding the following to your CMake
-DBOOST_ROOT:PATHNAME=$HOME/boost -DBoost_LIBRARY_DIRS:FILEPATH=$HOME/boost/lib -DBoost_NO_BOOST_CMAKE=TRUE -DBoost_NO_SYSTEM_PATHS=TRUE
Note that you will also have to set your
LD_LIBRARY_PATH to point to
Note also that Boost must be compiled with the same compiler version as
you are using to compile DyNet.
DyNet has been tested to build in Windows using Microsoft Visual Studio 2015. You may be able to build with MSVC 2013 by slightly modifying the instructions below.
First, install Eigen following the above instructions.
To generate the MSVC solution and project files, run cmake, pointing it to the location you installed Eigen and Boost (for example, at c:\libs\Eigen and c:\libs\boost_1_61_0):
mkdir build cd build cmake .. -DEIGEN3_INCLUDE_DIR=c:\libs\Eigen -G"Visual Studio 14 2015 Win64"
This will generate dynet.sln and a bunch of *.vcxproj files (one for the DyNet library, and one per example). You should be able to just open dynet.sln and build all. Note: multi-process functionality is currently not supported in Windows, so the multi-process examples (`*-mp`) will not be included in the generated solution
The Windows build also supports CUDA with the latest version of Eigen (as of Oct 28, 2016), with the following code change:
- TensorDeviceCuda.h: Change sleep(1) to Sleep(1000)