Introduction

This package aims to enable machine learning on AMR grid FLASH Code simulation data. FrankenCube uses PyTorch Lightning to setup a deep neural network for unsupervised learning.

Installation

Clone the FrankenCube repository

For installing the package you can either use the pip or conda package manager. It is recommended to use a virtual enviroment.

PIP

First create a virtual enviroment and activate it.

python -m venv <my-env>
source activate ./<my-env>/bin/activate

Afterwards install the requirements.

python -m pip install -r requirements.txt

CONDA

Create a virtual enviroment and install the dependencies in one line, afterwards activate your enviroment.

conda create --name <my-env> --file enviroment.yml
conda activate <my-env>

Setup FrankenCube

For our setup we take a look at /setup. Here you will find all of your setup files. At the the beginning you will see the subcube.yaml as the only setup file. The setup files handle almost everything for your machine learning runs.

To setup your first runs change the arguments inside the subcube.yaml file accordingly. Note that the package was intended to use AMR grid FLASH Code files, which may need to be preprocessed which /data/preprocessing.py.

Docker usage on Server

Sometimes a "normal" installation of PyTorch is not possible on a server and for that purpose containers are used. For that we need:

Prerequisites

Docker Engine
NVIDIA GPU Drivers
NVIDIA Container Toolkit

After installing the prerequisites we need to pick a NGC Container version and install everything with

docker run --gpus all -it --shm-size=4gb --name FrankenCube nvcr.io/nvidia/pytorch:24.03-py3

Quick Updates

Please make sure that if you are using a Docker container and want to install the package you may need to run:

apt update
apt upgrage

If you are using a virtual enviroment you may need to run:

apt-get install python3-venv

After which you can just run a virtualenviroment inside the container and install everything using the enviroments.txt.

Further reading

For further informations please refer to the Documentation and the Author

Some useful docker commands

Docker Cheat sheet

List all containers

docker ps -a

attach to a running docker container

docker attach <container image>

start a container

docker run [OPTIONS] <container image>

rename container

docker rename <old name> <new name>

save changes into new container

docker commit <containerID> <repository>:<tag>

Setup Github connection

First create a key pair for the Github Repo

ssh-keygen -t ed25519

The public key then gets attached to the Github under the settings segment of the Repo.

Add the private key to you ssh config file

Host FrankenCube
        Hostname github.com
        IdentityFile=~/.ssh/<YOUR PUB-KEYFILE>

Then clone the Repo

git clone git@FrankenCube:Tim-Boes/FrankenCube.git