agopal42 / ompn_craft

Introduction

This is the codebase of the paper Learning Task Decomposition with Order-Memory Policy Network It contains a version of the craft environment with gym wrappers. Dependency

python3.6
torch==1.5.1

Install locally pip install -e . If you find the environment or the paper to be useful, please cite

@inproceedings{
lu2021learning,
title={Learning Task Decomposition with Order-Memory Policy Network},
author={Yuchen Lu and Yikang Shen and Siyuan Zhou and Aaron Courville and Joshua B. Tenenbaum and Chuang Gan},
booktitle={International Conference on Learning Representations},
year={2021},
url={https://openreview.net/forum?id=vcopnwZ7bC}
}

Environments

import gym_psketch
print(gym_psketch.env_list)

Test

Keyboard interactive mode. Use arrow keys to move and u to use. Other key triggers done action.

python scripts/enjoy.py -mode keyboard -env [ENV_NAME]

See rule-based bot.

python scripts/enjoy.py -mode demo -env [ENV_NAME]

More scripts see in scripts

Train

Use main.py as main entry for both IL and RL

Imitation

Generate demo

python main.py --mode demo \
	--envs <ENVS_NAME> \
	--demo_episodes 1500

Run OMPN on unsupervised task information

python main.py --mode IL --arch omstack \
	--flagfile ilflagfile \ 
	--nb_slots 3 \
	--cuda \
	--envs <ENVS_NAME> \
	--env_arch noenv

Run OMPN on unsupervised with sketch information

python main.py --mode IL --arch omstack \
	--flagfile ilflagfile \ 
	--nb_slots 3 \
	--cuda \
	--envs <ENVS_NAME> \
	--env_arch sketch

Visualize the learned expanding position by

python scripts/analysis.py --model_ckpt PATH_TO_PKL \
        --envs makebedfull-v0 --use_demo --episodes 20