TensorFlow Probability (TFP) is a production-grade library for probabilistic modeling and Bayesian inference. It provides a comprehensive suite of probability distributions, bijectors (normalizing flows), MCMC kernels (HMC, NUTS, SMC, etc.), variational inference, Gaussian processes, structural time series, and Bayesian optimization. The library supports TensorFlow, JAX, and NumPy backends through a rewrite system, enabling seamless switching between hardware accelerators. The architecture is modular: each distribution, bijector, and kernel inherits from well-defined abstract base classes with `parameter_properties` for automatic shape and constraint inference. Key entry points include `tfp.distributions`, `tfp.bijectors`, `tfp.mcmc`, `tfp.vi`, `tfp.sts`, and `tfp.experimental` for cutting-edge methods. The code is heavily tested across all backends, with property-based tests and numerical stability checks. While primarily an application library, it also includes research-grade components in `discussion/` and `spinoffs/` (Inference Gym, AutoBNN, FunMC) that serve as benchmarks and experimental prototypes.
PyMC is a production-grade Bayesian modeling and probabilistic programming library built on PyTensor. It exposes a model-building API (pm.Model, pm.Normal, pm.sample), automatic log-probability derivation through graph rewriting, MCMC samplers (NUTS, HMC, Metropolis, Slice, DEMetropolis), Sequential Monte Carlo, variational inference (ADVI, FullRankADVI, SVGD), Gaussian processes, ODE support, and data containers (pm.Data, Minibatch). Users construct symbolic random-variable graphs using PyTensor; PyMC rewrites those graphs into measurable IR, computes log-probabilities and transforms, compiles sampling or variational objective functions, and returns posterior traces as xarray/ArviZ DataTree objects. Although the library primarily serves research and statistical workflows, the repository itself is an application/library, not a data-research pipeline. Architecturally, this is a library, not a service: there are no daemons or network endpoints. The entry points are Python APIs in pymc/model/core.py, pymc/sampling/mcmc.py, pymc/distributions/, pymc/step_methods/, and pymc/variational/. The core data flow is model specification -> PyTensor graph -> fgraph_from_model -> logprob rewrites -> compiled sampling functions -> trace backends (ndarray, zarr) -> to_inference_data for posterior analysis. External dependencies include NumPy, SciPy, PyTensor, ArviZ, and optional JAX/BlackJAX/NumPyro, Numba, Zarr, Rich, and Marimo. The codebase has an extensive test suite and CI, with some experimental areas such as pymc.dims and ongoing deprecations. Maintenance is active, with recent commits and rigorous CI. Technical drift is medium: most of the stack tracks current Python, but the Zarr backend is not yet ported to Zarr 3 and several optional integrations carry compatibility shims. Overall craftsmanship is strong but not exceptional: the code is deeply engineered, though its size and internal coupling make individual components less trivially extractable, and some deprecated APIs and experimental modules add friction.
Pyro is a deep probabilistic programming library built on PyTorch, providing a unified framework for Bayesian inference, variational inference, and Monte Carlo methods. Its architecture centers on composable effect handlers (poutine) that enable tracing, conditioning, replay, and reparameterization of probabilistic programs. The library includes a rich set of distributions, bijective transforms for normalizing flows, automatic guide construction, and a variety of inference algorithms (SVI, HMC, NUTS, SVGD, SMC). It also offers domain-specific contrib modules for Gaussian processes, time series forecasting, epidemiology, optimal experimental design, and multi-object tracking. While primarily a software library, Pyro has strong research roots, with many features inspired by recent papers and experimental modules that may evolve. The codebase is well-structured, extensively tested, and actively maintained, though some contrib modules carry instability warnings and a few internal patches to PyTorch may require attention.
probreg is a Python library designed for robust point cloud registration, implementing a suite of probabilistic algorithms including Coherent Point Drift (CPD), FilterReg, Bayesian CPD (BCPD), and GMMTree. It serves as a specialized tool for computer vision and robotics applications requiring the alignment of 3D point sets under rigid, affine, or non-rigid transformations. The library provides a clean Python API that abstracts complex mathematical operations, allowing users to register point clouds with support for both CPU (OpenMP) and GPU (CuPy) acceleration. Architecturally, the project follows a hybrid model where performance-critical kernels are implemented in C++ and exposed to Python via pybind11. This ensures high computational efficiency for operations like the Kabsch algorithm, point-to-plane distance calculations, and permutohedral lattice filtering. The Python layer manages the orchestration of these kernels, handling data I/O (via Open3D), transformation modeling, and iterative optimization loops. The codebase is well-structured with distinct modules for cost functions, transformations, and specific registration algorithms, promoting modularity and extensibility. The repository demonstrates strong engineering practices with a comprehensive test suite covering core algorithms, CI/CD pipelines for cross-platform builds (Ubuntu/Windows), and Sphinx documentation. It includes example scripts that serve as both documentation and validation for the library's functionality. The inclusion of C++ extensions and dependency on Eigen and pybind11 indicates a mature build system capable of handling complex native compilation requirements. The project is actively maintained with recent commits and a stable versioning strategy.
GPflow is a mature, pip-installable Python library for Gaussian process (GP) modeling built on TensorFlow and TensorFlow Probability. It is not a deployed service or CLI; it is a scientific software library that researchers and practitioners import directly. The library provides exact and approximate GP inference (GPR, SGPR, VGP, SVGP, GPMC, SGPMC, CGLB), a wide range of stationary, non-stationary, multi-output, convolutional, and hierarchical kernels, multiple likelihoods including quadrature and Monte Carlo fallbacks, and optimizers such as Scipy L-BFGS-B, natural gradient, and MCMC sampling helpers. The dominant architecture is modular and dispatch-driven: covariances, conditionals, expectations, KL divergences, and posterior construction are all implemented via a custom multiple-dispatch mechanism, making the library extensible to new kernel/inducing-variable/likelihood combinations. It also includes posterior caching for fast repeated predictions, TensorBoard monitoring, and shape annotations via check_shapes. While the repository contains extensive Jupyter/Sphinx notebooks and a dedicated benchmark harness, these are supporting assets rather than the primary deliverable; the core is clearly a reusable application library. The code is generally high-craftsmanship: clean abstractions, strong test coverage, and several genuinely generic components such as Gauss-Hermite quadrature helpers, a TensorFlow-module traversal utility, and a TF-Scipy optimizer wrapper with compile caching. However, the codebase shows meaningful drift risk: it supports a very wide TensorFlow/TFP version matrix (TF 2.4-2.16, TFP 0.12-0.24), uses private TensorFlow and TFP APIs in places, carries backward-compatibility/deprecation layers (e.g., GPR_deprecated, SVGP_with_posterior), and contains occasional debug leftovers or commented-out code. Maintenance is clearly active, with recent releases, CI matrices across Python 3.7-3.11, and versioned fixes to newly added hierarchical kernels.