This code provides a PyTorch implementation of the method described in this paper for training neural networks which can be used as proposal distributions for importance sampling or sequential Monte Carlo:
Paige, B., & Wood, F. (2016). Inference Networks for Sequential Monte Carlo in Graphical Models. In Proceedings of the 33rd International Conference on Machine Learning. JMLR W&CP 48: 3040-3049.
The largest section of re-usable code is an implementation of a conditional variant of MADE as a PyTorch module, found in learn_smc_proposals.cde.
This can be used to fit a conditional density estimator.
There is a version for real-valued data and a version for binary data.
- The linear regression notebook provides an end-to-end usage example. This notebook defines a generative model using PyMC, a non-conjugate regression model. It then goes through the process of defining a network which will represent the inverse, training it on samples from the joint distribution, and then using it for inference.
Two more involved examples are implemented in learn_smc_proposals.examples; pre-trained weights are included in this repository. Figures and inference are shown in two notebooks:
- A multilevel Poisson model demonstrates use of divide-and-conquer SMC for inference in a Bayesian heirarchical model;
- A factorial HMM demonstrates the binary data implementation and the repeated use of a single learned inverse in a particle filtering context.