uber/pyro
Pyro
Pyro is a probabilistic programming language and library built for PyTorch. It serves as a Bayesian inference engine and a tool for probabilistic graphical modeling, allowing users to define generative models that combine neural networks with probabilistic logic.
The framework enables deep probabilistic programming by integrating probability distributions into computational graphs. This allows for the quantification of uncertainty in deep learning models and the execution of scalable posterior distribution calculations for complex data dependencies.
The system provides a suite of inference capabilities, including variational inference and stochastic gradient Markov Chain Monte Carlo. It supports the construction of computable probability distributions and the estimation of latent variables through a tensor-based representation that leverages automatic differentiation.
Features
- Bayesian Probabilistic Programming Frameworks - Serves as a comprehensive Bayesian probabilistic programming framework for building models and performing inference.
- Posterior Inference - Implements mathematical routines for updating prior distributions with observed data to calculate posterior distributions.
- Probabilistic Graphical Models - Provides tools for representing complex dependencies in data using graph-based probability distributions.
- Probabilistic Models - Allows for the design of generative models and computable probability distributions to recover hidden patterns from data.
- Tensor Data Representations - Represents probability distributions using multi-dimensional tensors for hardware acceleration and automatic differentiation.
- Uncertainty Estimation - Integrates probabilistic logic into neural networks to provide uncertainty estimates and confidence intervals for predictions.
- Variational Inference - Provides an engine for approximating complex data likelihoods using latent variables and evidence lower bounds.
- Probabilistic Programming - Provides a complete framework for Bayesian inference, uncertainty quantification, and probabilistic modeling.
- Inference Execution - Provides scalable routines for executing Bayesian inference on complex probabilistic models.
- Inference Scaling - Enables scalable posterior distribution calculations and latent variable estimation for large datasets.
- Markov Chain Monte Carlo Sampling - Implements Markov Chain Monte Carlo sampling algorithms for estimating complex probability distributions.
- Automatic Guide Construction - Provides automated construction of guides to simplify the creation of variational inference approximations.
- Model - Records random sample calls during execution to enable deterministic playback during inference.
- Probabilistic and Generative - Library for deep universal probabilistic programming.
- Probabilistic Modeling - Deep probabilistic programming library built on PyTorch.
