Installing the Python DyNet module.¶
(for instructions on installing on a computer with GPU, see below)
Python bindings to DyNet are supported for both Python 2.x and 3.x.
TL;DR¶
(see below for the details)
# Installing Python DyNet:
pip install cython # if you don't have it already.
mkdir dynet-base
cd dynet-base
# getting dynet and eigen
git clone https://github.com/clab/dynet.git
hg clone https://bitbucket.org/eigen/eigen -r 346ecdb # -r NUM specified a known working revision
cd dynet
mkdir build
cd build
# without GPU support:
cmake .. -DEIGEN3_INCLUDE_DIR=../../eigen -DPYTHON=`which python`
# or with GPU support:
cmake .. -DEIGEN3_INCLUDE_DIR=../../eigen -DPYTHON=`which python` -DBACKEND=cuda
make -j 2 # replace 2 with the number of available cores
cd python
python setup.py install # or `python setup.py install --user` for a user-local install.
# this should suffice, but on some systems you may need to add the following line to your
# init files in order for the compiled .so files be accessible to Python.
# /path/to/dynet/build/dynet is the location in which libdynet.dylib resides.
export DYLD_LIBRARY_PATH=/path/to/dynet/build/dynet/:$DYLD_LIBRARY_PATH
Detailed Instructions¶
First, get DyNet:
cd $HOME
mkdir dynet-base
cd dynet-base
git clone https://github.com/clab/dynet.git
cd dynet
git submodule init # To be consistent with DyNet's installation instructions.
git submodule update # To be consistent with DyNet's installation instructions.
Then get Eigen:
cd $HOME
cd dynet-base
hg clone https://bitbucket.org/eigen/eigen/ -r 346ecdb
(-r NUM specifies a known working revision of Eigen. You can remove this in order to get the bleeding edge Eigen, with the risk of some compile breaks, and the possible benefit of added optimizations.)
We also need to make sure the cython
module is installed. (you can
replace pip
with your favorite package manager, such as conda
,
or install within a virtual environment)
pip install cython
To simplify the following steps, we can set a bash variable to hold where we have saved the main directories of DyNet and Eigen. In case you have gotten DyNet and Eigen differently from the instructions above and saved them in different location(s), these variables will be helpful:
PATH_TO_DYNET=$HOME/dynet-base/dynet/
PATH_TO_EIGEN=$HOME/dynet-base/eigen/
Compile DyNet.
This is pretty much the same process as compiling DyNet, with the
addition of the -DPYTHON=
flag, pointing to the location of your
Python interpreter.
If Boost is installed in a non-standard location, you should add the
corresponding flags to the cmake
commandline, see the DyNet
installation instructions page.
cd $PATH_TO_DYNET
PATH_TO_PYTHON=`which python`
mkdir build
cd build
cmake .. -DEIGEN3_INCLUDE_DIR=$PATH_TO_EIGEN -DPYTHON=$PATH_TO_PYTHON
make -j 2
Assuming that the cmake
command found all the needed libraries and
didn’t fail, the make
command will take a while, and compile DyNet
as well as the Python bindings. You can change make -j 2
to a higher
number, depending on the available cores you want to use while
compiling.
You now have a working Python binding inside of build/dynet
. To
verify this is working:
cd $PATH_TO_DYNET/build/python
python
then, within Python:
import dynet as dy
print dy.__version__
model = dy.Model()
In order to install the module so that it is accessible from everywhere in the system, run the following:
cd $PATH_TO_DYNET/build/python
python setup.py install --user
The --user
switch will install the module in your local
site-packages, and works without root privileges. To install the module
to the system site-packages (for all users), or to the current virtualenv
(if you are on one), run python setup.py install
without this switch.
You should now have a working python binding (the dynet
module).
Note however that the installation relies on the compiled DyNet library
being in $PATH_TO_DYNET/build/dynet
, so make sure not to move it
from there.
Now, check that everything works:
cd $PATH_TO_DYNET
cd examples/python
python xor.py
python rnnlm.py rnnlm.py
Alternatively, if the following script works for you, then your installation is likely to be working:
from dynet import *
model = Model()
If it doesn’t work and you get an error similar to the following:
ImportError: dlopen(/Users/sneharajana/.python-eggs/dyNET-0.0.0-py2.7-macosx-10.11-intel.egg-tmp/_dynet.so, 2): Library not loaded: @rpath/libdynet.dylib
Referenced from: /Users/sneharajana/.python-eggs/dyNET-0.0.0-py2.7-macosx-10.11-intel.egg-tmp/_dynet.so
Reason: image not found``
then you may need to run the following (and add it to your shell init files):
export DYLD_LIBRARY_PATH=/path/to/dynet/build/dynet/:$DYLD_LIBRARY_PATH
Usage¶
There are two ways to import the dynet module :
import dynet
imports dynet and automatically initializes the global dynet parameters with the command line arguments (see the documentation). The amount of memory allocated, GPU/CPU usage is fixed from there on.
import _dynet
# or
import _gdynet # For GPU
Imports dynet for CPU (resp. GPU) and doesn’t initialize the global parameters. These must be initialized manually before using dynet, using one of the following :
# Same as import dynet as dy
import _dynet as dy
dy.init()
# Same as import dynet as dy
import _dynet as dy
# Declare a DynetParams object
dyparams = dy.DynetParams()
# Fetch the command line arguments (optional)
dyparams.from_args()
# Set some parameters manualy (see the command line arguments documentation)
dyparams.set_mem(2048)
dyparams.set_random_seed(666)
dyparams.set_weight_decay(1e-7)
dyparams.set_shared_parameters(False)
dyparams.set_requested_gpus(1)
dyparams.set_gpu_mask([0,1,1,0])
# Initialize with the given parameters
dyparams.init() # or init_from_params(dyparams)
Anaconda Support¶
Anaconda is a popular package management system for Python. DyNet can be used from within an Anaconda environment, but be sure to activate the environment
source activate my_environment_name
then install some necessary packages as follows:
conda install gcc cmake boost cython
After this, the build process should be the same as normal.
Note that on some conda environments, people have reported build errors related to the interaction between the icu
and boost
packages. If you encounter this, try the solution in this comment.
Windows Support¶
You can also use Python on Windows by following similar steps to the above. For simplicity, we recommend using a Python distribution that already has Cython installed. The following has been tested to work:
- Install WinPython 2.7.10 (comes with Cython already installed).
- Run CMake as above with
-DPYTHON=/path/to/your/python.exe
. - Open a command prompt and set
VS90COMNTOOLS
to the path to your Visual Studio “Common7/Tools” directory. One easy way to do this is a command such as:
set VS90COMNTOOLS=%VS140COMNTOOLS%
- Open dynet.sln from this command prompt and build the “Release” version of the solution.
- Follow the rest of the instructions above for testing the build and installing it for other users
Note, currently only the Release version works.
GPU/MKL Support¶
Installing/running on GPU¶
For installing on a computer with GPU, first install CUDA. The following instructions assume CUDA is installed.
The installation process is pretty much the same, while adding the
-DBACKEND=cuda
flag to the cmake
stage:
cmake .. -DEIGEN3_INCLUDE_DIR=$PATH_TO_EIGEN -DPYTHON=$PATH_TO_PYTHON -DBACKEND=cuda
(if CUDA is installed in a non-standard location and cmake
cannot
find it, you can specify also
-DCUDA_TOOLKIT_ROOT_DIR=/path/to/cuda
.)
Now, build the Python modules (as above, we assume Cython is installed):
After running make -j 2
, you should have the files _dynet.so
and
_gdynet.so
in the build/python
folder.
As before, cd build/python
followed by
python setup.py install --user
will install the module.
In order to use the GPU support, you can either:
- Use
import _gdynet as dy
instead ofimport dynet as dy
- Or, (preferred),
import dynet
as usual, but use the commandline switch--dynet-gpu
or the GPU switches detailed here when invoking the program. This option lets the same code work with either the GPU or the CPU version depending on how it is invoked.
Running with MKL¶
If you’ve built DyNet to use MKL (using -DMKL
or -DMKL_ROOT
), Python sometimes has difficulty finding
the MKL shared libraries. You can try setting LD_LIBRARY_PATH
to point to your MKL library directory.
If that doesn’t work, try setting the following environment variable (supposing, for example,
your MKL libraries are located at /opt/intel/mkl/lib/intel64
):
export LD_PRELOAD=/opt/intel/mkl/lib/intel64/libmkl_def.so:/opt/intel/mkl/lib/intel64/libmkl_avx2.so:/opt/intel/mkl/lib/intel64/libmkl_core.so:/opt/intel/mkl/lib/intel64/libmkl_intel_lp64.so:/opt/intel/mkl/lib/intel64/libmkl_intel_thread.so:/opt/intel/lib/intel64_lin/libiomp5.so