Flexible Variational Information Bottleneck

The implementation codes of Flexible Variational Information Bottleneck: Achieving Diverse Compression with a Single Training.

To use in a project

See demo.ipynb for simple description of the usage.
The file utils.py contains functions for the learning.
The file fvib.py contains modules for FVIB and VIB.
The file loss.py contains loss functions for FVIB, VIB and the Taylor approximaition of the VIB objective.
The file calibration.py contains a module for continuous optimization of $\beta$ in FVIB and a module to calculate ECE.

Citation

If you find this library useful please consider citing our paper.