Vbq

[robamler]

Implement a fast Variational Bayesian Quantization with a dynamically adjusting empirical prior.

See:

• for the original proposal: Yang, Bamler, and Mandt, "Variatioinal Bayesian Quantization", ICML 2020
• for an evaluation on neural network compression: Tan and Bamler, "Post-Training Neural Network Compression With Variational Bayesian Quantization", Deploy & Monitor ML Workshop at NeurIPS 2022