Code for the paper "Information Condensing Active Learning" -- https://arxiv.org/abs/2002.07916
This code also uses code from the following papers
BatchBALD: Efficient and Diverse Batch Acquisition for Deep Bayesian Active Learning
Bayesian Batch Active Learning as Sparse Subset Approximation
and the repository
https://github.com/kuangliu/pytorch-cifar
Make sure you install all requirements using
conda install pytorch torchvision cudatoolkit=10.0 -c pytorch
pip install -r requirements.txt
The acquisition functions are in multi_bald.py. ICAL is implemented in the function compute_ical and ICAL-pointwise in compute_ical_pointwise