Build MXNet

outline of this page

set variables and paths
install the build dependencies
build and install the shared library
install the Python bindings
validate the installation
install the Perl bindings

The natural language projects presented on this website use MXNet and GluonNLP to predict word context for a Sicilian dictionary and to develop a machine translator.

And because I enjoy programming in Perl, the MXNet in Perl pages also introduce the AI::MXNet module by Sergey Kolychev, which he uses for computer vision projects.

MXNet provides a good set of installation instructions, but any good introduction starts with an installation, so I typed a few notes on how I built and installed MXNet on Debian Buster and how I built the API bindings for Python and Perl.

set variables and paths

After building the shared library, I wanted to store it in $HOME/.local/mxnet/lib, so my first step was to add the following variables and paths to my $HOME/.bashrc:

## local library path -- contains libmxnet.so and more

export LD_LIBRARY_PATH="$HOME/.local/mxnet/lib"

## python path

export PYTHONPATH="$HOME/.local/lib/python3.7/site-packages"

## perl path

export PERL5LIB="$HOME/.perl/lib/perl5"

## set PATH so it includes $HOME/.perl/bin

PATH="$HOME/.perl/bin:$PATH"

## set PATH so it includes $HOME/.local/bin

PATH="$HOME/.local/bin:$PATH"

## set PATH so it includes $HOME/.bin

PATH="$HOME/.bin:$PATH"

install the build dependencies

Next, I installed the build dependencies for the shared library:

# apt-get install build-essential ca-certificates cmake git liblapack-dev libmkldnn-dev libopenblas-dev libopencv-dev ninja-build python3-dev unzip wget

Python's pip3 installer:

# apt-get install python3-pip

and the dependencies for the Perl module:

# apt-get install libmouse-perl pdl cpanminus swig libgraphviz-perl

build and install the shared library

Having chosen those paths and variables and having installed the build dependencies, I downloaded the apache-mxnet-src-1.4.1-incubating.tar.gz tarball from GitHub and began to build.

First, I extracted the tarball and entered the directory:

$ tar -xzf apache-mxnet-src-1.4.1-incubating.tar.gz

$ cd apache-mxnet-src-1.4.1-incubating/

$ MXNET_HOME=${PWD}

Then, I built the shared library (saving a copy of the build log) using OpenCV for computer vision, OpenBLAS for linear algebra and Intel's Math Kernel Library for Deep Neural Networks (MKL-DNN):

$ make -j $(nproc) USE_OPENCV=1 USE_BLAS=openblas USE_MKLDNN=1

If you have a Nvidia graphics card, you might want to use the CUDA toolkit:

USE_CUDA=1 USE_CUDA_PATH=/usr/local/cuda USE_CUDNN=1

After building the shared libraries, I installed them by copying them to $HOME/.local/mxnet/lib

$ cp lib/libiomp5.so ${LD_LIBRARY_PATH}/.

$ cp lib/libmkldnn.so.0 ${LD_LIBRARY_PATH}/.

$ cp lib/libmklml_intel.so ${LD_LIBRARY_PATH}/.

$ cp lib/libmxnet.a ${LD_LIBRARY_PATH}/.

$ cp lib/libmxnet.so ${LD_LIBRARY_PATH}/.

install the Python bindings

Next, I installed the Python package:

$ cd ${MXNET_HOME}/python

$ pip3 install .

I also had to create a few symbolic links inside the installed package:

$ cd $HOME/.local/lib/python3.7/site-packages/mxnet

$ ln -s $HOME/.local/mxnet/lib/libmxnet.so libmxnet.so

$ ln -s $HOME/.local/mxnet/lib/libmklml_intel.so libmklml_intel.so

$ ln -s $HOME/.local/mxnet/lib/libiomp5.so libiomp5.so

validate the installation

After that, I validated the installation by passing the following commands to the python3 interpreter:

>>> import mxnet as mx

>>> a = mx.nd.ones((2, 3))

>>> b = a * 2 + 1

>>> b.asnumpy()

which returned:

array([[3., 3., 3.],

[3., 3., 3.]], dtype=float32)

which shows that MXNet and the Python bindings were successfully installed.

To test MKL-DNN, I also performed the validation at the MKL-DNN ReadMe page.