GradBED

Research code for the submission "Gradient-Based Bayesian Experimental Design for Implicit Models using Mutual Information Lower Bounds". This repository is actively being updated to make the code more usable for others.

CPU Setup

Install conda dependencies and the project with

conda env create -f environment.yml
conda activate gradbed-env
python setup.py develop

Separately install the torchsde package for simulating the SDE-based epidemiological models:

pip install git+https://github.com/google-research/torchsde.git

If the dependencies in environment.yml change, update dependencies with

conda env update --file environment.yml

GPU Cluster Setup

Check local versions of cuda available: ls -d /opt/cu*. You should use one of these (e.g. the latest version) for the cudatoolkit=??.? argument below.

Create a Conda environment with GPU-enabled PyTorch (with e.g. Cuda 10.1):

conda create -n gradbed-env python=3.8 pytorch torchvision cudatoolkit=10.1 -c pytorch
conda activate gradbed-env

Then install dependencies in the GPU environment file:

conda env update --file environment-gpu.yml

Separately install the torchsde package for simulating the SDE-based epidemiological models:

pip install git+https://github.com/google-research/torchsde.git

The above command with the environment file can also be used to update the Conda environment when dependencies in the environment file change.