Installing Analysis-G

Installing via Github

First clone the project into a target directory.

git clone

To automate most of the setup, navigate to the setup-scripts directory and run the script. This will generate a new Python environment called GNN, which can be sourced from the shell script

cd AnalysisG/setup-scripts && bash

If you are running the framework on a HPC cluster, for instance lxplus or some other Linux environment, make sure to have at least GCC 6.20 enabled. For example, on lxplus machines, add the following lines to the bashrc file:

export ATLAS_LOCAL_ROOT_BASE=/cvmfs/
source ${ATLAS_LOCAL_ROOT_BASE}/user/
lsetup "gcc gcc620_x86_64_slc6"

If making changes to the .bashrc is not desired, simply add an alias to it and point the source command to the setup-scripts folder as shown below:

alias GNN='source <some path to repository>/setup-scripts/

This will create a command called GNN and when executed will setup the Analysis environment.

Installing via PyPI (pip)

The framework has also been published to PyPI to simplify the installation process.

pip install analysisg

Additional Software Setup

Analysis-G is partly dependent on PyC and can be installed via the command install_pyc. Unlike most PyTorch packages, the installation process is rather seemless. During the build process, the package will scan for nvcc (CUDA compiler), and will install the appropriate PyTorch version (CUDA/CPU).

As mentioned in the introduction page of these docs, modules found in this package are completely written in C++. If CUDA is available, then the package will also proceed to install the native CUDA kernel implementations. This however can be a very long and computationally expensive build process.

Unlike the recommended setup-tools based tutorial provided by PyTorch, the installation process utilizes an advanced cmake builder called scikit-build-core. This allows for code modifications to be made without having to repetitively recompile unmodified code and wasting computational resources. Once installed, the module can be used via:

import pyc

Install via Command-Line

cd torch-extensions
pip install -v .

Install via Framework