This is a fork of the CRAFT repo that is used to evaluate the CRAFT model on the GMM and Many Well problems for our paper Flow Annealed Importance Sampling Bootstrap. We show that our method (FAB) significantly outperforms CRAFT on the GMM problem - CRAFT fails to capture all the modes in the target distribution on this problem. On the Many Well problem both FAB and CRAFT perform well, where both methods capture the modes in the target distribution well.
Install notes: make sure to use chex==0.1.5
Colab notebook for running experiment for GMM problem:
Notebook for running experiment for Many Well Problem
These experiments may also be run using the appropriate configs inside main.py.
