Deep Adaptive Design: Amortizing Sequential Bayesian Experimental Design

An unofficial implementation of the ICML 2021 paper: Deep Adaptive Design: Amortizing Sequential Bayesian Experimental Design by Foster et al.¹

Structure

• config/: hydra configuration files
• data/: simulate experiments
• loss/: loss functions
• model/: design policy networks
• utils/: helper functions

Setup

Requirements:

• Python ≥ 3.10
• PyTorch ≥ 2.0
• Additional dependencies are listed in requirements.txt

Installation:

Create a new virtual environment:

# create a new conda environment
conda create -n dad python=3.12 -y
# activate the environment
conda activate dad

Install PyTorch (adjust based on your OS and CUDA version, see: https://pytorch.org/get-started/locally/):

pip install torch==2.6.0 --index-url https://download.pytorch.org/whl/cu126

Install remaining dependencies:

pip install -r requirements.txt

Experiments

To reproduce the location-finding experiment with default settings:

python location_finding.py

To run with custom hyperparameters:

python location_finding.py data.K=1 data.theta_dist=uniform

Logging with Weights & Biases

To enable logging and monitor training in real time, add the following flag:

wandb.use_wandb=True

Make sure you are logged into your Weights & Biases account.

Credits

This repository aims to reproduce the main experimental results from the paper with simplified dependencies and a modular structure. If you find it useful, please feel free to build on it. Contributions are warmly welcome and appreciated :)

For the official implementation, please refer to dad.

Footnotes

1. Foster, Adam, Desi R. Ivanova, Ilyas Malik, and Tom Rainforth. ‘Deep Adaptive Design: Amortizing Sequential Bayesian Experimental Design’. In Proceedings of the 38th International Conference on Machine Learning, 3384–95. PMLR, 2021. https://proceedings.mlr.press/v139/foster21a. ↩