Fast ε-free Inference of Simulation Models with Bayesian Conditional Density Estimation

Code for reproducing the experiments in the paper:

G. Papamakarios and I. Murray, Fast ε-free Inference of Simulation Models with Bayesian Conditional Density Estimation, NeurIPS 2016. [arXiv] [bibtex]

This folder contains

demos

Folder containing four subfolders, one for each demo in the paper.

• mixture_of_gaussians_demo

  • mog_main.py --- sets up the model
  • mog_abc.py --- runs ABC methods
  • mog_mdn.py --- runs MDN methods
  • mog_res.py --- collects and plots results

• bayesian_linear_regression_demo

  • blr_main.py --- sets up the model
  • blr_abc.py --- runs ABC methods
  • blr_mdn.py --- runs MDN methods
  • blr_res.py --- collects and plots results

• lotka_volterra_demo

  • lv_main.py --- sets up the model
  • lv_abc.py --- runs ABC methods
  • lv_mdn.py --- runs MDN methods
  • lv_res.py --- collects and plots results

• mg1_queue_demo

  • mg1_main.py --- sets up the model
  • mg1_abc.py --- runs ABC methods
  • mg1_mdn.py --- runs MDN methods
  • mg1_res.py --- collects and plots results

util

Folder with utility classes and functions.

• pdf.py Gaussians and mixtures of Gaussians

• NeuralNet.py neural nets with and without SVI

• mdn.py MDNs with and without SVI

• DataStream.py provides data minibatches for training

• LossFunction.py loss functions for training

• StepStrategy.py optimization algorithms, including Adam

• Trainer.py trains a neural net or MDN, SVI or not

• MarkovJumpProcess.py Markov jump processes, including Lotka--Volterra

• helper.py various helper functions