https://github.com/lxuechen/bdmc

PyTorch implementation of Bidirectional Monte Carlo, Annealed Importance Sampling, and Hamiltonian Monte Carlo.
https://github.com/lxuechen/bdmc

annealed-importance-sampling bayesian-inference bidirectional-monte-carlo hamiltonian-monte-carlo mcmc pytorch

Last synced: 20 days ago
JSON representation

PyTorch implementation of Bidirectional Monte Carlo, Annealed Importance Sampling, and Hamiltonian Monte Carlo.

• Host: GitHub
• URL: https://github.com/lxuechen/bdmc
• Owner: lxuechen
• License: mit
• Created: 2018-01-06T20:45:10.000Z (almost 7 years ago)
• Default Branch: master
• Last Pushed: 2021-05-14T21:39:59.000Z (over 3 years ago)
• Last Synced: 2024-10-14T11:09:50.603Z (about 1 month ago)
• Topics: annealed-importance-sampling, bayesian-inference, bidirectional-monte-carlo, hamiltonian-monte-carlo, mcmc, pytorch
• Language: Python
• Homepage:
• Size: 3.05 MB
• Stars: 52
• Watchers: 3
• Forks: 14
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# BDMC

PyTorch implementation of Bidirectional Monte Carlo.

## Requirements

* `python3`

* `numpy`

* `pytorch>=0.4.1`

* `tqdm`

## What is Bidirectional Monte Carlo (BDMC)?

BDMC is a method of accurately sandwiching the log marginal likelihood (ML). It is mainly used to evaluate the quality

of log-ML estimators [1]. The method achieves this by obtaining a lower bound with the usual Annealed Importance

Sampling (AIS) [2], and an upper bound with Reversed AIS from an exact posterior sample. Since the upper bound requires

an *exact* sample from the posterior, the method is only strictly valid on simulated data. However, the results obtained

on simulated data can help verify the performance of log-ML estimators. Conditioned upon the assumption that the real

data does not differ too much from the simulated data, the evaluation of the log-ML estimator on simulated data could be

informative of the performance on real data.

The given implementation performs evaluation on a variational autoencoder (VAE) trained on MNIST.

## To run

There is a pretrained VAE model (on MNIST) in the `checkpoints` folder. Executing the command

```bash

python bdmc.py \

  --latent-dim 50 \

  --batch-size 512 \

  --n-batch 2 \

  --chain-length 10000 \

  --iwae-samples 10 \

  --ckpt-path ./checkpoints/model.pth

```

will start the forward and backwards chain of BDMC based on the model loaded from the pretrained checkpoint.

## References

[1] Grosse, Roger B., Zoubin Ghahramani, and Ryan P. Adams. "Sandwiching the marginal likelihood using bidirectional

Monte Carlo." arXiv preprint arXiv:1511.02543 (2015).

[2] Neal, Radford M. "Annealed importance sampling." Statistics and computing 11.2 (2001): 125-139.

[3] Neal, Radford M. "MCMC using Hamiltonian dynamics." Handbook of Markov Chain Monte Carlo 2.11 (2011).