goalNCA

Codebase for training Neural Cellular Automata (NCAs) to grow, morph, rotate, and translate target patterns conditioned on task vectors.

Main directories:

• NCAs/: NCA model implementations (MLP-based and LSTM-based variants).
• src/trainers/: Training loop and multiplexing experiment orchestration.
• src/datasets/: Dataset classes for pattern loading and transformation tasks.
• src/visualisation/: Visualization and animation utilities.
• src/utils/: General utility functions and plotting helpers.
• src/experiments/: Multiplexing experiment scripts.
• datasets/: Raw dataset files (emoji/pattern images).

---

Installation

Install dependencies using uv and the provided pyproject.toml or uv.lock.

First install uv, then simply type uv sync.

---

Usage

Training

To train an NCA, run:

uv run python train.py --conf config_growing.yml

Supported tasks:

• Conditional Growth: Grow patterns from seeds conditioned on task vectors.
• Morphing: Morph from one pattern to another.
• Translation: Translate patterns in space.

Available configuration files:

• config_growing.yml: Conditional growth of patterns from seeds.
• config_morphing.yml: Morphing between patterns.
• config_translation.yml: Translating patterns in space.

Results are saved to results/{task_name}/{run_name}/.

Evaluation

To evaluate a trained model:

uv run python evaluate.py --path results/{task_name}/{run_name}/

---

Configuration

Configuration files (config_*.yml) control all aspects of training, including:

• Task type and target patterns (emoji unicode, image paths, or dataset folders)
• Model architecture (layers, dimensions, activations, convolution mode)
• NCA dynamics (steps, noise, stochastic updates, boundary conditions)
• WandB experiment tracking