Growing Neural Cellular Automata

This repository is reimplementation on PyTorch code from this blog

We can formulate task as using self-organizing model of cellular automata to reconstruct predefined pattern from any state.

Example of output from trained model:

All experiments was on running Ubuntu OS, NVDIA 2080 TI GPU.

Requirements:

* Python 3.6+
* **16GB+ RAM Memory**
* CUDA 9.1+ (For GPU training)
* 10GB GPU Memory (For GPU training) 

Steps to reproduce:

1. Run pip install -r requirements.txt.
2. Run python main.py to use CPU or python main.py --use-cuda to use GPU
3. Wait while training ends, in sonic/infer_log folder will be generated images

Make GIF

0. (Optional) Run for resize in <images_folder> folder for X in *; do convert $X -interpolate Nearest -filter point -resize 480x480 $X; done
1. To make GIF run convert -delay 20 -loop 1 <images_folder>/*.jpg myimage.gif

How to decrease Memory usage:

• For GPU(with --use-cuda parameter) or CPU(w/o --use-cuda parameter) you can change next parameters in config:

  1. decrease both values in ITER_NUMBER tuple
  2. decrease BATCH_SIZE
  3. decrease TARGET_SIZE

• For CPU only usage:

  1. decrease POOL_SIZE