Fast Neural Machine Translation in C++

Перейти к файлу

Lane Schwartz e851f31dd4 Exclude 3rd party code from generated docs		2016-10-03 12:31:54 -05:00
examples	merge	2016-09-16 11:48:03 +02:00
marian	eclipse	2016-09-22 18:39:36 +01:00
scripts	avoid reallocation of temporary tensors	2016-09-21 02:38:12 +02:00
src	Merge branch 'master' of github.com:emjotde/Marian	2016-10-03 11:34:45 -05:00
.gitignore	add functions loading MNIST dataset	2016-09-13 18:13:23 +02:00
CMakeLists.txt	removed dependencies on extended lambdas, fixes #15	2016-09-22 15:16:03 +02:00
Doxyfile.in	Exclude 3rd party code from generated docs	2016-10-03 12:31:54 -05:00
LICENSE.md	Update LICENSE.md	2016-09-30 00:13:14 +01:00
README.md	Added MNist doc	2016-09-21 12:24:19 -05:00

README.md

Marian

Google group for commit messages: https://groups.google.com/forum/#!forum/mariannmt

A C++ gpu-specific parallel automatic differentiation library with operator overloading.

In honour of Marian Rejewski, a Polish mathematician and cryptologist.

Installation

Requirements:

g++ with c++11
CUDA and CuDNN
Boost (>= 1.56)

Exporting some paths for CuDNN may be required (put it, for example, in your .bashrc file):

export PATH=$PATH:$HOME/.local/bin:/usr/local/cuda/bin
export LIBRARY_PATH=$LIBRARY_PATH:/usr/local/cuda/lib64:/usr/local/cuda/lib64/stubs:/usr/local/cudnn-5/lib64
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda/lib64:/usr/local/cuda/lib64/stubs:/usr/local/cudnn-5/lib64
export CPATH=$CPATH:/usr/local/cudnn-5/include

Compilation with cmake > 3.5:

mkdir build
cd build
cmake ..
make -j

To compile API documentation using Doxygen, first cd to the build directory, and then:

make doc

To test, first compile, then:

cd examples/mnist
make
cd ../../build
./mnist_benchmark