Neural-Cellular-Automata

Goals

Our goals can be broken down into answering three topics:

1. What are NCA? How is NCA different from other NCA?
2. What can NCA be used for? Does NCA provide an advantage over other similar architectures?
3. How can NCA be improved?

As a result of answering these questions, we aim to produce a research paper.

Open Source!

All of our code for our website and for generating website content is open source can be found in this repository! There are two main parts:

1. The Training Code: This produces the weights an biases of our neural network. All implemented in PyTorch. There will be multiple versions of this for various experiments. There is a secondary repository for our 3D work which can be found 3D Neural Cellular Automata
2. The Website: Models are rendered in WebGPU and TypeScript. The website itself is built in Next.js. Output of this code is the website.

Resources

This project is currently using the Notion platform to document project progress and important information. This Notion workspace will be made public at a later point in time.

However, some useful resources for this project include:

• Understanding Cellular Automata (CA)
  • Introduction to Conway's Game of Life
  • Explaining CA
  • What are "Life-Like" CAs
• Neural Cellular Automata
  • Neural Patterns
  • Growing Neural Cellular Automata
• Tutorials for Building CAs
  • Daniel Shiffman's Tutorial
  • Building Conway's Game of Life using WebGPU
  • Physics Simulation with CA
• Other
  • Noita uses CA to make physics simulation
  • "Rule-String" Notation
  • CAs and Computational Systems
  • Emergent Gardens
• WebGPU Resources
  • WebGPU Shader Tips
  • Typescript & WebGPU Examples

More resources will be made available in this repo as the project progresses.

Repository Maintenance

We will maintain at most 3 types of branches on the repository for this project. How each branch type will be used is detailed in dev cycles. Details on each branch type are elaborated here:

Main Branch

The branch that is always operational. It contains features that have been tested, are fully functional, and are ready to be seen by the public. As such, this branch is to be treated extremely carefully. Main will only branch to dev, and only dev will merge to main. The reason for this is so we can vet and check everything before anything goes up to main. Pushing straight to main should only be done never.

Dev Branch

This is the branch that hold all things under development. Every time a feature is finished, it will be merged to dev. On dev we then have the chance to do bug fixing and benchmarking. We do all the necessary test to ensure the stuff we upload is genuinely ready.

Feature Branches

These branches are where actual programming occurs. Create a feature branch from main, and name it accordingly. Try to keep it down to one feature per brach, so the git history will clearly outline our development process.

There is a lot of freedom in what you are allowed to do in this branch type, since everything will be vetted when you merge to dev, and more vetting will be done before merging to main. The principle is just try to keep things reasonable and clear.

Dev Cycles

This project will be done in "dev cycles". That's just an outline how we will conduct development and keep everything organised. The general progression of a dev cycle is as follows.

1. Branch a new dev branch from main.
2. Create feature branches from dev, and add your new features.
3. Upon completion, feature branches merge back to dev.
4. Bug testing on dev. These may occur as a result of feature integration.
5. After bug testing (and more benchmarking maybe) is complete, Then Dev should be merged back to main.
6. Dev is archived, and the Dev cycle is complete.

Setup

This project works on Python 3.12 and may also work on versions newer than Python 3.9. To install and manage multiple versions of Python, we recommend using miniconda. This is especially useful, if you plan on using Python for other projects.

Conda Environment

To create a new environment for this project, run the following commands in your terminal (assuming you have a GPU that supports CUDA 12.4):

conda create -n nca --file requirements.txt

If you do not have a discrete GPU, or have an older GPU that does not support CUDA 12.4, you can create an environment manually with the following commands:

conda create -n nca python==3.12
conda activate nca
pip install matplotlib

You will also need to install the correct PyTorch version with conda using their installation guide. Ensure that you choose Conda as the package manager, and select the correct CUDA version for your GPU or CPU.

Using the Environment

To use the environment, run the following command in your terminal:

conda activate nca

You can also automatically activate the environment for this project in VSCode by selecting the correct Python interpreter in the bottom right corner of the window (when a Python file is open).