# metabeta

**Repository Path**: Tomhappy/metabeta

## Basic Information

- **Project Name**: metabeta
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-02-08
- **Last Updated**: 2026-02-08

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# metabeta
---
🚧 I'm currently refactoring this repo. For legacy code, switch to branch v1. 🚧
---
## A fast neural model for Bayesian mixed-effects regression
📈 approximates posterior estimation for hierarchical regression models\
💡 allows prior specification for each regression parameter\
👨🏻‍🏫 trained on simulated hierarchical datasets with realistic structure\
🚀 orders of magnitude faster than HCMC\
🔥 fully implemented in [PyTorch](https://pytorch.org/)

This repo contains the source code for [data simulation](metabeta/simulation), [model architecture](metabeta/models), and [evaluation](metabeta/evaluation).

## Overview
1. Hierarchical datasets are summarized locally (per group) and globally (across groups).
2. During training, a normalizing flow learns the forward mapping (regression parameters → base distribution) for both parameter types.
3. During inference, samples are drawn from the base distribution, and are mapped backward (regression parameters ← base distribution).

## Setup with [uv](https://docs.astral.sh/uv/)
- uv install
- source .venv/bin/activate
- uv pip install -e .