Growing cellular automata (as described in https://distill.pub/2020/growing-ca/) adapted to 3D objects.
To reproduce everything, setup a new environment using python 3.8 and then install all dependencies
python -m pip install -r requirements.txt
To run the streamlit demo:
streamlit run streamlit-demo.py
Animals in PLY format downloaded from SketchFab-WaxFreeOintment
Each model is trained for 3000-6000 steps (higher is the amount of pixels, more steps will require to converge) using Colab GPUs. With a Nvidia T4 (Colab) the time required may vary from 30 minutes to 1 hour using a batch size of 4 and a pool size of 128.
For each animal object there are 3 checkpoints:
- distillWay which is the strategy reproduced from GNCA distill paper, adapted to 3D objects. To reproduce this, just set modifiedTrainingMode to False in DataModuleConfig.
- modifiedNoNoiseWay which consists in using a cropped/cut-out damage at the spot of the simple type of damage. To reproduce this, set modifiedTrainingMode to True and randomNoise to False.
- modifiedWay same as modifiedNoNoiseWay but using also a random substitution of cells with random values to let the model persist/recover even in very strange states. To reproduce this, set modifiedTrainingMode and randomNoise to True.
