Awesome

BDI

We propose BiDirectional learning for offline Infinite-width model-based optimization (BDI) between the high-scoring designs and the static dataset (a.k.a. low-scoring designs).

If you find this code useful in your research then please cite:

@inproceedings{
zhang2022bidirectional,
title={Bidirectional Learning for Offline Infinite-width Model-based Optimization},
author={Can Chen and Yingxue Zhang and Jie Fu and Xue Liu and Mark Coates},
booktitle={Thirty-Sixth Conference on Neural Information Processing Systems},
year={2022},
url={https://openreview.net/forum?id=_j8yVIyp27Q}
}

Installation

The environment of BDI can be installed as:

conda create --name BDI --file requirements.txt
conda activate BDI

Reproducing Performance

For the TF Bind 8 task, we can run the forward mapping as:

python -u BDI.py --mode grad --task TFBind8-Exact-v0 --outer_lr 1e-1 --gamma 0.0

The backward mapping can be run as:

python -u BDI.py --mode distill --task TFBind8-Exact-v0 --outer_lr 1e-1 --gamma 0.0

Run our BDI as:

python -u BDI.py --mode both --task TFBind8-Exact-v0 --outer_lr 1e-1 --gamma 0.0

Similarly for AntMorphology task, we have:

python -u BDI.py --mode grad --task AntMorphology-Exact-v0 --outer_lr 1e-3 --gamma 0.001
python -u BDI.py --mode distill --task AntMorphology-Exact-v0 --outer_lr 1e-3 --gamma 0.001
python -u BDI.py --mode both --task AntMorphology-Exact-v0 --outer_lr 1e-3 --gamma 0.001

Acknowledgements

We thank the design-bench library (https://github.com/brandontrabucco/design-bench) and the data distillation implementation (https://colab.research.google.com/github/google-research/google-research/blob/master/kip/KIP.ipynb).