Repository for my masters thesis that studies stochastic variational inference for semiparametric distributional regression models. Variational inference is a inference approach that approximates a posterior distribution using a simpler variational distribution. This inference technique allows to learn a posterior distribution with optimization instead of sampling. The thesis studies a "black-box" variational inference algorithm that can be applied to a wide variety of hierarchical Bayesian regression models.
The directory tigerpy includes python code of the masters thesis that is loosely written in a package style. The python package is called tigerpy and allows for Bayesian inference in distributional regression models that can also contain smooth additive effects (via the B-spline basis). Flexible regression specifications are enabled by its directed graph structure. The package networkx is used for efficient directed graph construction. Moreover makes the package use of jax for high-performance array computing and automatic differentiation.
The master thesis is written in quarto and is fully contained in the thesis directory. Child chapters contained in the chapters directory are all imported into the master file thesis_paper.qmd and knitted there.
A simulation study is contained in the simulation directory. While the playground directory contains .ipynb files that apply the inference algorithm to different models. Additionally some figures for the thesis are generated in the notebook plots_paper.ipynb.
The variational inference algorithm is based on Ranganath et al. 2014 and Kucukelbir et al. 2016. A well written general introduction into variational inference is provided by Zhang et al. 2017.
.
├── README.md
├── playground
├── simulation
├── thesis
│ ├── bib
│ ├── chapters
│ └── tex
└── tigerpy
├── bbvi
├── distributions
└── model
Some literature references that might provide good background when studying the repo.
- Advances in Variational Inference
- Variational Inference: A Review for Statisticians
- Black Box Variational Inference
- Automatic Differentiation Variational Inference
- Liesel
Use the pre-commit system library.
Try to follow the Google Python Style Guide.