PyTorch implementation for the ICLR 2020 paper Understanding the Limitations of Variational Mutual Information Estimators.

by Jiaming Song and Stefano Ermon, Stanford Artificial Intelligence Laboratory.

The code depends on PyTorch >= 1.2, numpy, pandas and matplotlib. It has been tested on both Python 3.7.

We implement several mutual information estimators, including:

• InfoNCE: Contrastive predictive coding / Info Noise Contrastive Estimation.
• NWJ: Variational representation of the KL divergence (lower bound).
• NWJ (JS): Train with variational representation of JS divergence lower bound, evaluate with KL.
• MINE / DV: Variational representation of the KL divergence based on Donsker-Varadhan inequality.
• SMILE: our method with clipping for estimating partition functions.

These functions are implemented in estimators.py.

See demo.ipynb for the procedures to produce the figures in the paper.

If you use this code for your research, please cite our paper:

@article{song2020understanding,
  title="Understanding the Limitations of Variational Mutual Information Estimators",
  author="Song, Jiaming and Ermon, Stefano",
  conference="International Conference on Learning Representations",
  year="2020"
}

tsong@cs.stanford.edu