ensmallen-graph 0.6.0

No project description provided

Python Bindings

How to install this

This project is currently work in progress, and is to be considered for all intents and porposes an alpha version.

To install the latest (alpha) release, run the following:

pip install ensmallen_graph

Compile the bindings

The project uses PyO3 and maturin. You need nightly rust in order to compile PyO3. This could be done by running:

curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh --default-toolchain nightly --profile default -y

Then we need to install maturin with:

cargo install maturin

And finally we can create the Wheels with:

maturin build --release

The wheels will be located in ./target/wheels/.

Compile the bindings for multiple Python versions

By default maturin will build for the installed python version.

To install other Python versions just run:

wget https://www.python.org/ftp/python/${PYTHON_VERSION}/Python-${PYTHON_VERSION}.tar.xz
tar -xf Python-${PYTHON_VERSION}.tar.xz
(cd Python-${PYTHON_VERSION}; ./configure --enable-optimizations; make -j$(nproc); make -j$(nproc) install)

Where ${PYTHON_VERSION} is the choosen version, such as 3.8.3 or 3.6.10.

Solving the GLIBC__2.29__ not found error

This means that the linux package were built on a system with a really new Libc version. This can be solved by compiling your own bindings. In order to be compatible with as many systems as possible, we suggest to build them on the official Manylinux1 docker.

Exploiting Avx

The library vec_rand which is used in the crate can take advantage in using Avx instructions/

To compile the bindings with Avx Instructions you can run :

RUSTFLAGS=" -C target-cpu=native" maturin build --release --no-sdist

Compiling the bindings for Linux to push them to Pipy

In order to make the wheels usable by any linux system we must follow the manylinux2010 standard. We used to support manylinux1 but rust no longer support Glibc 2.5 so we had to bump it to the next standard.

This can be done by building the bindings inside a centos5 docker with:

git clone https://github.com/LucaCappelletti95/ensmallen_graph
cd ensmallen_graph
make build_manylinux2010 python_manylinux2010

The result wheels will be in ./ensmallen_graph/bindings/python/target/wheels/.

A wheel will be created for each version of python, this is configurable in the Dockerfile ./ensmallen_graph/setup/DockerFileManylinux2010 in a ENV var.

A tutorial on how the internals of Python's Cffi can be found here