An approximate sampler for energy-based models with divergence diagnostics

Quasi Rejection Sampling (QRS) is a new method for sampling from energy-based models with explicit diagnostics for computing divergence. A tunable parameter allows to trade-off fidelity of the sampling distribution for sampling efficiency, allowing to find the best practical trade-off for any given application.

For more information, check out the paper published in TMLR - November 2022.

To get access to the code, check out the QRS GitHub repository.