pxpy 1.0a71

discrete pairwise undirected graphical models

Copyright (c) 2021 Nico Piatkowski

pxpy

The python library for discrete pairwise undirected graphical models. Runs on Linux with GLIBC >= 2.28 and Windows 10. Exported ONNX models run on any architecture that has an ONNX runtime with opset 13.

Inference

• Loopy belief propagation

• Junction tree

• Stochastic Clenshaw-Curtis quadrature

Sampling

• Gibbs Sampling

• Perturb+Map Sampling

Parameter learning

• Accelerated proximal gradient

• built-in L1 / L2 regularization

• Support for custom regularization

Structure learning

• Chow-Liu trees

• Soft-thresolding

• High-order clique structures

Misc

• Support for deep Boltzmann tree models (DBT)

• Support for spatio-temporal compressible reparametrization (STRF)

• Runs on x86_64 (linux, windows) and aarch64 (linux)

• Graph drawing via graphviz

• Discretization

<https://randomfields.org>

—

Alpha Changelog

• 1.0a71: Improved: ONNX and QUBO export

• 1.0a70: Added: QUBO export

• 1.0a69: Added: ONNX PAM export

• 1.0a68: Improved: Data type selection, ONNX export

• 1.0a67: Improved: Integer MRF; ONNX export

• 1.0a66: Improved: Data type selection, various minor fixes. Added: Integer model file io.

• 1.0a65: Improved: Manual model construction

• 1.0a64: Improved: GPU code. Added: ONNX export

• 1.0a63: Added: Experimental annealed rejection sampler for structure sampling

• 1.0a62: Improved: Model loading

• 1.0a61: Improved: Setting target for “star” structure; reduced python version to 3.6

• 1.0a60: Improved: Numerical stability of discretization

• 1.0a55: Added: Load/store of discretization models; aarch64 support (tested on Jetson TX1)

• 1.0a54: Improved: Init speed

• 1.0a53: Improved: Init speed

• 1.0a52: Improved: Graph splitting; init speed

• 1.0a51: Fixed: Multi-core normalization; Split-edge weight centering

• 1.0a50: Improved: Support for external inference engines; Changed required GLIBC version to 2.29

• 1.0a49: Fixed: External loader

• 1.0a48: Added: Shell script “pxpy_environ” for populating various environment variables. Improved: multi-core support.

• 1.0a47: Added: draw_neighbors(..). Improved: Discretization

• 1.0a44: Improved: Discretization

• 1.0a42: Improved: Updated some default values

• 1.0a41: Improved: Fixed subtle bug in parameter initialization

• 1.0a40: Added: Loading string data via genfromstrcsv(..) (built-in string<->int mapper)

• 1.0a36: Improved: Randomized clique search

• 1.0a29: Added: Randomized clique search

• 1.0a28: Improved: Handling NaN-values during discretization (now interpreted as missing)

• 1.0a27: Improved: Accelerated structure estimation

• 1.0a26: Improved: Progress computation. Added: Online entropy computation for large cliques

• 1.0a25: Improved: Memory management

• 1.0a24: Improved: Structure estimation, backend. Added: Third-order structure estimation; simple graphviz output

• 1.0a23: Improved: Structure estimation

• 1.0a22: Improved: Discretization engine, support for external inference engine. Added: default to 32bit computation (disable via env PX_USE64BIT)

• 1.0a21: Improved: Support for external inference engine

• 1.0a20: Added: Support for external inference engine (access via env PX_EXTINF)

• 1.0a19: Improved: Manual model creation

• 1.0a18: Added: Debug mode (linux only, enable via env PX_DEBUGMODE)

• 1.0a17: Improved: API, tests, regularization. Added: AIC and BIC computation

• 1.0a16: Improved: Memory management, access to optimizer state in optimization hooks. Added: Support for training resumption

• 1.0a15: Improved: API

• 1.0a14: Improved: Memory management

• 1.0a13: Improved: Memory management (fixed leak in conditional sampling/marginals)

• 1.0a12: Improved: Access to vertex and pairwise marginals

• 1.0a11: Added: Access to single variable marginals

• 1.0a10: Improved: Library build process

• 1.0a9: Added: Conditional sampling

• 1.0a8: Imroved: Maximum-a-posteriori (MAP) estimation. Added: Custom graph construction

• 1.0a7: Added: Conditional marginal inference, support for Ising/minimal statistics

• 1.0a6: Added: Manual model creation, support for training data with missing values (represented by pxpy.MISSING_VALUE)

• 1.0a5: Improved: Model management

• 1.0a4: Added: Model access in regularization and proximal hooks

• 1.0a3: Improved: GLIBC requirement, removed libgomp dependency

• 1.0a2: Added: Python 3.5 compatibility

• 1.0a1: Initial release