Refining Diffusion Planner for Reliable Behavior Synthesis by Automatic Detection of Infeasible Plans
This repository provides the source codes for our paper Refining Diffusion Planner for Reliable Behavior Synthesis by Automatic Detection of Infeasible Plans in NeurIPS 2023.
pip install -e .
pip install captum==0.6.0 gtimer
cd third_party/diffuser
pip install -e .We offer pretrained models for gap prediction as well as pre-collected synthetic data. To download and set them up, run the following commands:
bash prepare_pretrained_models.sh
bash prepare_data.shExecuting these commands will download and extract tar files, placing their contents into the logs/gap_predictor and logs/data directories respectively.
Evaluate RGG:
python scripts/maze2d/evaluate_rgg.py \
--logbase ./logs/evaluate \
--env_name maze2d-large-v1 \
--task single_task \
--diffusers_repo leekwoon/maze2d-large-v1-H384-T256 \
--num_episodes 1000 \
--spec rgg \
--gap_predictor_path ./logs/pretrained_gap_predictor/maze2d-large-v1-H384-T256/500000_finish/2023_03_26_03_44_06/seed_0/state_best.ptEvaluate RGG+:
python scripts/maze2d/evaluate_rggplus.py \
--logbase ./logs/evaluate \
--env_name maze2d-large-v1 \
--task single_task \
--diffusers_repo leekwoon/maze2d-large-v1-H384-T256 \
--num_episodes 1000 \
--spec rggplus \
--gap_predictor_path ./logs/pretrained_gap_predictor/maze2d-large-v1-H384-T256/500000_finish/2023_03_26_03_44_06/seed_0/state_best.pt- Generate synthetic data with:
python scripts/maze2d/make_data.py \
--base_datadir ./logs/data \
--env_name maze2d-large-v1 \
--diffusers_repo leekwoon/maze2d-large-v1-H384-T256 \
--n 500000- Compute restoration gap with:
python scripts/maze2d/compute_restoration_gaps.py \
--env_name maze2d-large-v1 \
--data_path ./logs/data/maze2d-large-v1-H384-T256/500000_finish.npz \
--diffusers_repo leekwoon/maze2d-large-v1-H384-T256 \
--strength 0.9 \
--num_plan 10- Train a gap predictor with:
python scripts/train_gap_predictor.py \
--base_logdir ./logs/gap_predictor \
--env_name maze2d-large-v1 \
--data_path ./logs/data/maze2d-large-v1-H384-T256/500000_finish.npz \
--score_path ./logs/data/maze2d-large-v1-H384-T256/500000_finish_restoration_gaps.npy \
--seed 0Attribution map can be inspected using notebooks/maze2d/plot_attribution.ipynb.
@inproceedings{lee2023refining,
title={Refining Diffusion Planner for Reliable Behavior Synthesis by Automatic Detection of Infeasible Plans},
author={Lee, Kyowoon and Kim, Seongun and Choi, Jaesik},
booktitle={Advances in Neural Information Processing Systems},
year={2023},
}
This repository is released under the MIT license. See LICENSE for additional details.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent