Contrastive Learning of Structured World Models

Last update: Jan 06, 2023

Related tags

Overview

Contrastive Learning of Structured World Models

This repository contains the official PyTorch implementation of:

Contrastive Learning of Structured World Models.
Thomas Kipf, Elise van der Pol, Max Welling.
http://arxiv.org/abs/1911.12247

Abstract: A structured understanding of our world in terms of objects, relations, and hierarchies is an important component of human cognition. Learning such a structured world model from raw sensory data remains a challenge. As a step towards this goal, we introduce Contrastively-trained Structured World Models (C-SWMs). C-SWMs utilize a contrastive approach for representation learning in environments with compositional structure. We structure each state embedding as a set of object representations and their relations, modeled by a graph neural network. This allows objects to be discovered from raw pixel observations without direct supervision as part of the learning process. We evaluate C-SWMs on compositional environments involving multiple interacting objects that can be manipulated independently by an agent, simple Atari games, and a multi-object physics simulation. Our experiments demonstrate that C-SWMs can overcome limitations of models based on pixel reconstruction and outperform typical representatives of this model class in highly structured environments, while learning interpretable object-based representations.

Requirements

Python 3.6 or 3.7
PyTorch version 1.2
OpenAI Gym version: 0.12.0 pip install gym==0.12.0
OpenAI Atari_py version: 0.1.4: pip install atari-py==0.1.4
Scikit-image version 0.15.0 pip install scikit-image==0.15.0
Matplotlib version 3.0.2 pip install matplotlib==3.0.2

Generate datasets

2D Shapes:

python data_gen/env.py --env_id ShapesTrain-v0 --fname data/shapes_train.h5 --num_episodes 1000 --seed 1
python data_gen/env.py --env_id ShapesEval-v0 --fname data/shapes_eval.h5 --num_episodes 10000 --seed 2

3D Cubes:

python data_gen/env.py --env_id CubesTrain-v0 --fname data/cubes_train.h5 --num_episodes 1000 --seed 3
python data_gen/env.py --env_id CubesEval-v0 --fname data/cubes_eval.h5 --num_episodes 10000 --seed 4

Atari Pong:

python data_gen/env.py --env_id PongDeterministic-v4 --fname data/pong_train.h5 --num_episodes 1000 --atari --seed 1
python data_gen/env.py --env_id PongDeterministic-v4 --fname data/pong_eval.h5 --num_episodes 100 --atari --seed 2

Space Invaders:

python data_gen/env.py --env_id SpaceInvadersDeterministic-v4 --fname data/spaceinvaders_train.h5 --num_episodes 1000 --atari --seed 1
python data_gen/env.py --env_id SpaceInvadersDeterministic-v4 --fname data/spaceinvaders_eval.h5 --num_episodes 100 --atari --seed 2

3-Body Gravitational Physics:

python data_gen/physics.py --num-episodes 5000 --fname data/balls_train.h5 --seed 1
python data_gen/physics.py --num-episodes 1000 --fname data/balls_eval.h5 --eval --seed 2

Run model training and evaluation

2D Shapes:

python train.py --dataset data/shapes_train.h5 --encoder small --name shapes
python eval.py --dataset data/shapes_eval.h5 --save-folder checkpoints/shapes --num-steps 1

3D Cubes:

python train.py --dataset data/cubes_train.h5 --encoder large --name cubes
python eval.py --dataset data/cubes_eval.h5 --save-folder checkpoints/cubes --num-steps 1

Atari Pong:

python train.py --dataset data/pong_train.h5 --encoder medium --embedding-dim 4 --action-dim 6 --num-objects 3 --copy-action --epochs 200 --name pong
python eval.py --dataset data/pong_eval.h5 --save-folder checkpoints/pong --num-steps 1

Space Invaders:

python train.py --dataset data/spaceinvaders_train.h5 --encoder medium --embedding-dim 4 --action-dim 6 --num-objects 3 --copy-action --epochs 200 --name spaceinvaders
python eval.py --dataset data/spaceinvaders_eval.h5 --save-folder checkpoints/spaceinvaders --num-steps 1

3-Body Gravitational Physics:

python train.py --dataset data/balls_train.h5 --encoder medium --embedding-dim 4 --num-objects 3 --ignore-action --name balls
python eval.py --dataset data/balls_eval.h5 --save-folder checkpoints/balls --num-steps 1

Cite

If you make use of this code in your own work, please cite our paper:

@article{kipf2019contrastive,
  title={Contrastive Learning of Structured World Models}, 
  author={Kipf, Thomas and van der Pol, Elise and Welling, Max}, 
  journal={arXiv preprint arXiv:1911.12247}, 
  year={2019} 
}

Contrastive Learning of Structured World Models

Related tags

Overview

Contrastive Learning of Structured World Models

Requirements

Generate datasets

Run model training and evaluation

Cite

Owner

Thomas Kipf

The official implementation of the Hybrid Self-Attention NEAT algorithm

Python utility to generate filesystem content for Obsidian.

On the adaptation of recurrent neural networks for system identification

A clean and scalable template to kickstart your deep learning project 🚀 ⚡ 🔥

ICS 4u HD project, start before-wards. A curtain shooting game using python.

Hypersearch weight debugging and losses tutorial

A python interface for training Reinforcement Learning bots to battle on pokemon showdown

This is the official PyTorch implementation of the paper "TransFG: A Transformer Architecture for Fine-grained Recognition" (Ju He, Jie-Neng Chen, Shuai Liu, Adam Kortylewski, Cheng Yang, Yutong Bai, Changhu Wang, Alan Yuille).

[CVPR 2022] Official Pytorch code for OW-DETR: Open-world Detection Transformer

Hyper-parameter optimization for sklearn

A torch.Tensor-like DataFrame library supporting multiple execution runtimes and Arrow as a common memory format

Certis - Certis, A High-Quality Backtesting Engine

A Low Complexity Speech Enhancement Framework for Full-Band Audio (48kHz) based on Deep Filtering.

Code for "Continuous-Time Meta-Learning with Forward Mode Differentiation" (ICLR 2022)

Code for Dual Contrastive Learning for Unsupervised Image-to-Image Translation, NTIRE, CVPRW 2021.

Gray Zone Assessment

Kaggle-titanic - A tutorial for Kaggle's Titanic: Machine Learning from Disaster competition. Demonstrates basic data munging, analysis, and visualization techniques. Shows examples of supervised machine learning techniques.

I explore rock vs. mine prediction using a SONAR dataset

NLG evaluation via Statistical Measures of Similarity: BaryScore, DepthScore, InfoLM

Code for "Modeling Indirect Illumination for Inverse Rendering", CVPR 2022