Unsupervised Discovery of Parts, Structure, and Dynamics

This repository contains pre-trained models and testing code for PSD presented at ICLR 2019.

http://psd.csail.mit.edu

Installation

Our current release has been tested on Ubuntu 18.04 LTS.

Cloning the repository

git clone git@github.com:zhenjia-xu/psd.git
cd psd

Set up Python environment:

pipenv install --dev

Guide

Generating shape dataset

pipenv run python data_generator.py

Downloading the pretrained model (185 MB)

./download_model.sh

Visualization

pipenv run python demo.py --gpu ID

• --gpu ID: use which gpu (starting from 0). Set to -1 to use CPU only.

The results will be presented in ./demo/index.html.

Reference

@inproceedings{psd,
  title={Unsupervised Discovery of Parts, Structure, and Dynamics},
  author={Xu, Zhenjia and Liu, Zhijian and Sun, Chen and Murphy, Kevin and
        Freeman, William T and Tenenbaum, Joshua B and Wu, Jiajun},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2019},
}

For any questions, please contact Zhenjia Xu (xuzhenjia@cs.columbia.edu).