Loopy belief propagation for factor graphs on discrete variables, in JAX!

Related tags

Deep Learningpythonjax
Overview

continuous-integration PyPI version pre-commit.ci status codecov Documentation Status

PGMax

PGMax implements general factor graphs for discrete probabilistic graphical models (PGMs), and hardware-accelerated differentiable loopy belief propagation (LBP) in JAX.

  • General factor graphs: PGMax supports easy specification of general factor graphs with potentially complicated topology, factor definitions, and discrete variables with a varying number of states.
  • LBP in JAX: PGMax generates pure JAX functions implementing LBP for a given factor graph. The generated pure JAX functions run on modern accelerators (GPU/TPU), work with JAX transformations (e.g. vmap for processing batches of models/samples, grad for differentiating through the LBP iterative process), and can be easily used as part of a larger end-to-end differentiable system.

See our blog post and companion paper for more details.

Installation | Getting started

Installation

Install from PyPI

pip install pgmax

Install latest version from GitHub

pip install git+https://github.com/vicariousinc/PGMax.git

Developer

git clone https://github.com/vicariousinc/PGMax.git
curl -sSL https://raw.githubusercontent.com/python-poetry/poetry/master/install-poetry.py | python3 -
cd PGMax
poetry shell
poetry install
pre-commit install

Install on GPU

By default the above commands install JAX for CPU. If you have access to a GPU, follow the official instructions here to install JAX for GPU.

Getting Started

Here are a few self-contained Colab notebooks to help you get started on using PGMax:

Citing PGMax

Please consider citing our companion paper if you use PGMax in your work:

@article{zhou2022pgmax,
  author = {Zhou, Guangyao and Kumar, Nishanth and L{\'a}zaro-Gredilla, Miguel and Kushagra, Shrinu and George, Dileep},
  title = {{PGMax: Factor Graphs for Discrete Probabilistic Graphical Models and Loopy Belief Propagation in JAX}},
  journal = {arXiv preprint arXiv:2202.04110},
  year={2022}
}

First two authors contributed equally.

Comments
  • Incomplete documentation for add_factor

    Incomplete documentation for add_factor

    The documentation for add_factor currently says:

    log_potentials – Optional array of shape (num_val_configs,) or (num_factors, num_val_configs). If specified, it contains the log of the potential value for every possible configuration. If none, it is assumed the log potential is uniform 0 and such an array is automatically initialized.

    However, in order to use it one would need to know the order in which the factors should be specified. Could this be added to the documentation?

    bug documentation 
    opened by nathanielvirgo 8
  • Add support for Python 3.9 and 3.10

    Add support for Python 3.9 and 3.10

    I don't know if this is a bug or a problem with my installation, but if I try to run a file containing only the line

    from pgmax.fg import graph

    I get the error

    % /opt/local/bin/python3 /Users/nathaniel/Dropbox/Code/PGMax/test01.py
Traceback (most recent call last):
  File "/Users/nathaniel/Dropbox/Code/PGMax/test01.py", line 1, in <module>
    from pgmax.fg import graph
  File "/Users/nathaniel/Library/Python/3.9/lib/python/site-packages/pgmax/fg/graph.py", line 11, in <module>
    import pgmax.bp.infer as infer
  File "/Users/nathaniel/Library/Python/3.9/lib/python/site-packages/pgmax/bp/infer.py", line 6, in <module>
    import pgmax.bp.bp_utils as bp_utils
  File "/Users/nathaniel/Library/Python/3.9/lib/python/site-packages/pgmax/bp/bp_utils.py", line 11, in <module>
    @jax.partial(jax.jit, static_argnames="max_segment_length")
AttributeError: module 'jax' has no attribute 'partial'

    This is with Python3.9 installed using Macports on MacOS 10.15.7. PGMax, jax and other prerequisities were installed with pip-3.9 install --user PGMax. I'm happy to give any other information about my installation if you can tell me how to obtain it.

    enhancement 
    opened by nathanielvirgo 6
  • Test sanity check example using new interface and inference modules, and put together the first unit test

    Test sanity check example using new interface and inference modules, and put together the first unit test

    The unit test should run fast. One option is to cache new results. Another option is to just make the model really small.

    In the process, we should also:

    1. Deprecate the current contrib module and create a new examples directory to hold everything.
    2. Start figuring out what our user facing interface should look like.
    opened by StannisZhou 6
  • FactorGraph supports any type of factors + runs specialized inference for ORFactors

    FactorGraph supports any type of factors + runs specialized inference for ORFactors

    In this PR we

    1. Redefine the factor graph abstraction by introducing factor types: factors in a graph are clustered in factor groups, which are grouped according to their factor types. See fg/graph.py
    2. Specify two types of factors: EnumerationFactor (this class already existed) and ORFactor (this new class inherits from the new LogicalFactor). Each Factor class must have its own methods to compile and concatenate Wirings for inference. See factors/enumeration.py and factors/logical.py.
    3. Make running inference in a graph agnostic to the current type of factors supported. New factors types can then be added without modifying graph.py.
    4. Implement a specialized inference for ORFactors (see pass_OR_fac_to_var_messages in factors/logical.py) and compare it with the existing one for EnumerationFactors in the unit test tests/factors/test_or.py
    opened by antoine-dedieu 5
  • RCN example

    RCN example

    This PR contains an example implementation of RCN using the pgmax package. We load a pre-trained RCN model on a very small subset of mnist (20 examples) and test on a small subset of mnist (20 examples). The reported accuracy = 0.80.

    opened by shrinuKushagra 5
  • Variables refactor

    Variables refactor

    We update the way of representing variables. In particular:

    • We get rid of variables names, as welll as of the Variables and CompositeVariableGroup classes. A variable is now represented by a tuple (variable hash, variable num_states) In particular, a FactorGraph can then directly be instantiated asfg = graph.FactorGraph(variables=[hidden_variables, visible_variables]) Similarly, Factors are defined by directly passing the variables involved, as [hidden_variables[ii], visible_variables[jj]]
    • We rewrite NDVariableArray so that the user can access variables by relying on the use of numpy arrays. We also optimize some follow-up computations.
    opened by antoine-dedieu 4
  • Numba speedup for wiring + log potentials

    Numba speedup for wiring + log potentials

    This PR is the continuation of https://github.com/vicariousinc/PGMax/pull/129 and part of our efforts to speed up the adding of FactorGroups and the wiring compilation.

    As https://github.com/vicariousinc/PGMax/pull/129 has moved most of the wiring computation to the FactorGroup level, we can now use numba for fast computation of these wirings

    As a result:

    • adding factors for the RBM exp takes 3s, building run_bp takes 1s
    • adding factors for the convor exp takes 2s, building run_bp takes 1s
    opened by antoine-dedieu 4
  • RCN implementation on a small train and test set

    RCN implementation on a small train and test set

    This PR contains the first implementation of the RCN example using the PGMax package. The file to run is examples/rcn/inference_pgmax_small.py This code contains implementation, visualization on a small set. Trained with 20 examples and tested on 20 examples.

    The inference has been separated from model creation code. Saved models are added to /storage/users/skushagra/pgmax_rcn_artifacts/ .

    Implementation on the full dataset will be implemented in a later PR.

    opened by shrinuKushagra 4
  • Make `FactorGraph` mutable to support interactive model building

    Make `FactorGraph` mutable to support interactive model building

    Should implement interface for:

    1. Add factors
    2. Set evidence for variables
    3. Initialize messages by setting messages for factors
    4. Initialize messages by spreading beliefs from variables
    enhancement 
    opened by StannisZhou 4
  • Add customized class for pairwise factors; Default to have uniform potentials

    Add customized class for pairwise factors; Default to have uniform potentials

    Currently, users have to manually create an array of all possible configs and a uniform potential, but it would be nice to do this behind-the-scenes in some easy way. Maybe we can make it so that if either of these is None during init, then we assume all possible configs or uniform potential respectively and automatically create these

    enhancement 
    opened by NishanthJKumar 4
  • Make BP closer to jax optimizer

    Make BP closer to jax optimizer

    Resolves https://github.com/vicariousinc/PGMax/issues/124

    We make graph.BP closer to JAX optimizers https://jax.readthedocs.io/en/latest/jax.example_libraries.optimizers.html

    opened by antoine-dedieu 3
  • [pre-commit.ci] pre-commit autoupdate

    [pre-commit.ci] pre-commit autoupdate

    opened by pre-commit-ci[bot] 1
  • Provide high-level syntax for creating factors

    Provide high-level syntax for creating factors

    One of the speed bottleneck in creating a FactorGraph is the time to create the variables_for_factors list, which is currently slow as we loop through the individual variables.

    However, in the case where all the variable groups are NDVarArr we can speed up this step a lot proposing a generic get_factors interface where the user would define the general rule for the factors and the corresponding list would be generated with numba.

    One options is to have a first argument which consists of variable groups for which we loop over dimensions, and a second argument which consists of variable groups for do not loop over, For, example get_factors({x:(i, j), y:(k, l)}, {z:(i+k, j+l)}) would mean

    factors = []
for i in range(x.shape[0]):
    for j in range(x.shape[1]):
        for k in range(y.shape[0]):
            for l in range(y.shape[1]): 
                factors.append((x[i, j], y[k, l], z[i+k, j+l]))
    opened by antoine-dedieu 0
  • Modify `vars_to_starts` representation

    Modify `vars_to_starts` representation

    Creating the vars_to_starts as a dictionnary mapping variable to int is expensive in the case where we have a lot of variables.

    Instead it could map a variable group to an array (in the case of a NDVariableArray) or a list (for a VariableDict)

    opened by antoine-dedieu 0
  • Improve documentation for variables/variable groups

    Improve documentation for variables/variable groups

    Currently it's not clear what names PGMax assigns to different variables (e.g. https://github.com/vicariousinc/PGMax/issues/115). Add documentation to make this clearer.

    documentation 
    opened by StannisZhou 0
Releases(v0.4.1)

  • v0.4.1(May 19, 2022)

    Highlights

    • Fixing two minor issues when running BP with variable groups defined with different number of states by @antoine-dedieu in https://github.com/vicariousinc/PGMax/pull/144

    Full Changelog: https://github.com/vicariousinc/PGMax/compare/v0.4.0...v0.4.1

    Source code(tar.gz)
    Source code(zip)

  • v0.4.0(May 9, 2022)

    Breaking changes

    ⚠️ This release changes the high-level API as well as import paths ⚠️

    This release makes several major breaking changes to improve usability and efficiency of the package.

    1. Interacting with variables through VarGroup objects

    We no longer refer to variables by names, but instead directly interact with VarGroup objects. This change has several implications.

    • We no longer have a Variable class. Instead, we access individual variables by indexing into VarGroup objects.

    • FactorGraph can no longer be initialized with a dictionary of variable groups (as we no longer have names for variables). Instead, we initialize a FactorGraph by

    from pgmax import fgraph
fg = fgraph.FactorGraph(variable_groups=variable_groups)

    where variable_groups is either a VarGroup or a list of VarGroups.

    • We can directly construct Factor/FactorGroup using individual variables, and have a unified add_factors interface for adding Factors and FactorGroups to the FactorGraph.

    For example, we can create a PairwiseFactorGroup via:

    from pgmax import fgroup
pairwise_factors = fgroup.PairwiseFactorGroup(
    variables_for_factors=variables_for_factors,
    log_potential_matrix=log_potential_matrix,
)

    where variables_for_factors is a list of list of individual variables. And we can add factors to a FactorGraph fg by

    fg.add_factors(factors=factors)

    where factors can be individual Factor, individual FactorGroup, or a list of Factors and FactorGroups.

    • We access LBP results by indexing with VarGroup. For example, after running BP, we can get the MAP decoding for the VarGroup visible_variables via
    beliefs = bp.get_beliefs(bp_arrays)
map_states_visible = infer.decode_map_states(beliefs)[visible_variables]

    2. Efficient construction of FactorGroup

    We have implemented efficient construction of FactorGroup. Going forward, we always recommend constructing FactorGroup instead of individual Factor.

    3. Improved LBP interface

    We first create the functions used to run BP with temperature T via

    from pgmax import infer
bp = infer.BP(fg.bp_state, temperature=T)

    where bp contains functions that initialize or updates the arrays involved in LBP.

    We can initialize bp_arrays by

    bp_arrays = bp.init()

    apply log potentials, messages and evidence updates by

    bp_arrays = bp.update(
    bp_arrays=bp_arrays,
	log_potentials_updates=log_potentials_updates,
	ftov_msgs_updates=ftov_msgs_updates,
	evidence_updates=evidence_updates,
)

    and run bp for a certain number of iterations by

    bp_arrays = bp.run_bp(bp_arrays, num_iters=num_iters, damping=damping)

    Note that we can arbitrarily interleave bp.update with bp.run_bp, which allows flexible control over how we run LBP.

    4. Improved high-level module organization

    Now we have 5 main high-level modules, fgraph for factor graphs, factor for factors, vgroup for variable groups, fgroup for factor groups, and infer for LBP.

    Details of what has changed:

    • Speed up the process of adding Factors and compiling wiring for a FactorGraph by moving all the computations to the FactorGroup level, by @antoine-dedieu in https://github.com/vicariousinc/PGMax/pull/129
    • Speed up the process of computing log potentials + wiring for FactorGroup with numba, by @antoine-dedieu in https://github.com/vicariousinc/PGMax/pull/133
    • Make the BP class behavior closer to JAX optimizers by @antoine-dedieu in https://github.com/vicariousinc/PGMax/pull/135
    • Get rid of the Variables and CompositeVariableGroup classes + of the variable names + adopt a simpler representation for variables + rely on numpy arrays to makeNDVarArray efficient, by @antoine-dedieu in https://github.com/vicariousinc/PGMax/pull/136
    • Overall module reorganization, by @antoine-dedieu in https://github.com/vicariousinc/PGMax/pull/140

    Full Changelog: https://github.com/vicariousinc/PGMax/compare/v0.3.0...v0.4.0

    Source code(tar.gz)
    Source code(zip)

  • v0.3.0(Mar 25, 2022)

    Highlights

    • Refactors to support adding different factor types with specialized inference procedures by @antoine-dedieu in https://github.com/vicariousinc/PGMax/pull/122
    • Specialized logical AND/OR factors by @antoine-dedieu in https://github.com/vicariousinc/PGMax/pull/122 https://github.com/vicariousinc/PGMax/pull/126
    • New example on 2D binary blind deconvolution by @antoine-dedieu in https://github.com/vicariousinc/PGMax/pull/127

    New Contributors

    • @antoine-dedieu made his first contribution in https://github.com/vicariousinc/PGMax/pull/122

    Full Changelog: https://github.com/vicariousinc/PGMax/compare/v0.2.3...v0.3.0

    Source code(tar.gz)
    Source code(zip)

  • v0.2.3(Feb 19, 2022)

    What's Changed

    • Links to blog post and companion paper; Documentation updates by @StannisZhou in https://github.com/vicariousinc/PGMax/pull/111
    • Get rid of redundant array shape for log_potentials by @StannisZhou in https://github.com/vicariousinc/PGMax/pull/116
    • Support python 3.9/3.10; Improve documentation for add_factor; Bump up version for new release by @StannisZhou in https://github.com/vicariousinc/PGMax/pull/120

    Full Changelog: https://github.com/vicariousinc/PGMax/compare/v0.2.2...v0.2.3

    Source code(tar.gz)
    Source code(zip)

  • v0.2.2(Jan 22, 2022)

    What's Changed

    • Update README by @StannisZhou in https://github.com/vicariousinc/PGMax/pull/103
    • RCN example by @shrinuKushagra in https://github.com/vicariousinc/PGMax/pull/96
    • Add support for sum-product with temperature by @StannisZhou in https://github.com/vicariousinc/PGMax/pull/104
    • Include Grid Markov Random Field example by @StannisZhou in https://github.com/vicariousinc/PGMax/pull/107
    • Changes for blog post by @StannisZhou in https://github.com/vicariousinc/PGMax/pull/109

    New Contributors

    • @shrinuKushagra made their first contribution in https://github.com/vicariousinc/PGMax/pull/96

    Full Changelog: https://github.com/vicariousinc/PGMax/compare/v0.2.1...v0.2.2

    Source code(tar.gz)
    Source code(zip)

  • v0.2.1(Dec 1, 2021)

    What's Changed

    • Bump versions for publishing by @StannisZhou in https://github.com/vicariousinc/PGMax/pull/63
    • Example notebook with PMAP sampling of RBMs trained on MNIST digits by @StannisZhou in https://github.com/vicariousinc/PGMax/pull/80
    • Use functools.partial instead of jax.partial by @StannisZhou in https://github.com/vicariousinc/PGMax/pull/83
    • First pass for speeding up graph and evidence operations by @StannisZhou in https://github.com/vicariousinc/PGMax/pull/84
    • Moving to a functional interface by @StannisZhou in https://github.com/vicariousinc/PGMax/pull/88
    • Update README in preparation for making repo public by @StannisZhou in https://github.com/vicariousinc/PGMax/pull/89
    • Pre commit ci test by @NishanthJKumar in https://github.com/vicariousinc/PGMax/pull/90
    • [pre-commit.ci] pre-commit autoupdate by @pre-commit-ci in https://github.com/vicariousinc/PGMax/pull/91
    • Update dependency requirements to be less aggresive by @StannisZhou in https://github.com/vicariousinc/PGMax/pull/92
    • adds codecov badge to README! by @NishanthJKumar in https://github.com/vicariousinc/PGMax/pull/94
    • Fix GPU memory leak that came up in RCN example by @StannisZhou in https://github.com/vicariousinc/PGMax/pull/97
    • [pre-commit.ci] pre-commit autoupdate by @pre-commit-ci in https://github.com/vicariousinc/PGMax/pull/95
    • Docs update by @NishanthJKumar in https://github.com/vicariousinc/PGMax/pull/98
    • fixes bug where wrong conf.py path was specified by @NishanthJKumar in https://github.com/vicariousinc/PGMax/pull/99
    • Rtd warning fix by @NishanthJKumar in https://github.com/vicariousinc/PGMax/pull/100
    • includes minor changes to README and documentation by @NishanthJKumar in https://github.com/vicariousinc/PGMax/pull/101
    • Bump up version; Fixes for docs by @StannisZhou in https://github.com/vicariousinc/PGMax/pull/102

    Full Changelog: https://github.com/vicariousinc/PGMax/compare/v0.2.0...v0.2.1

    Source code(tar.gz)
    Source code(zip)

  • v0.2.0(Sep 6, 2021)

    Features

    • Efficient and scalable max-product belief propagation using a fully flat representation
    • A general factor graph interface that supports easy specification of PGMs with pairwise factors and higher-order factors based on explicit enumeration
    • Mechanisms for evidence and message manipulation
    • 3 example notebooks showcasing the functionalities of PGMax
    Source code(tar.gz)
    Source code(zip)
Owner
Vicarious
Vicarious
GitHub Repository https://pgmax.readthedocs.io
A resource for learning about ML, DL, PyTorch and TensorFlow. Feedback always appreciated :)

A resource for learning about ML, DL, PyTorch and TensorFlow. Feedback always appreciated :)

Aladdin Persson 4.7k Jan 08, 2023
Pytorch implementation for the EMNLP 2020 (Findings) paper: Connecting the Dots: A Knowledgeable Path Generator for Commonsense Question Answering

Path-Generator-QA This is a Pytorch implementation for the EMNLP 2020 (Findings) paper: Connecting the Dots: A Knowledgeable Path Generator for Common

Peifeng Wang 33 Dec 05, 2022
Urban mobility simulations with Python3, RLlib (Deep Reinforcement Learning) and Mesa (Agent-based modeling)

Deep Reinforcement Learning for Smart Cities Documentation RLlib: https://docs.ray.io/en/master/rllib.html Mesa: https://mesa.readthedocs.io/en/stable

1 May 15, 2022
MaRS - a recursive filtering framework that allows for truly modular multi-sensor integration

The Modular and Robust State-Estimation Framework, or short, MaRS, is a recursive filtering framework that allows for truly modular multi-sensor integration

Control of Networked Systems - University of Klagenfurt 143 Dec 29, 2022
Unsupervised captioning - Code for Unsupervised Image Captioning

Unsupervised Image Captioning by Yang Feng, Lin Ma, Wei Liu, and Jiebo Luo Introduction Most image captioning models are trained using paired image-se

Yang Feng 207 Dec 24, 2022
This program uses trial auth token of Azure Cognitive Services to do speech synthesis for you.

🗣️ aspeak A simple text-to-speech client using azure TTS API(trial). 😆 TL;DR: This program uses trial auth token of Azure Cognitive Services to do s

Levi Zim 359 Jan 05, 2023
Clean and readable code for Decision Transformer: Reinforcement Learning via Sequence Modeling

Minimal implementation of Decision Transformer: Reinforcement Learning via Sequence Modeling in PyTorch for mujoco control tasks in OpenAI gym

Nikhil Barhate 104 Jan 06, 2023
Final project code: Implementing BicycleGAN, for CIS680 FA21 at University of Pennsylvania

680 Final Project: BicycleGAN Haoran Tang Instructions 1. Training To train the network, please run train.py. Change hyper-parameters and folder paths

Haoran Tang 0 Apr 22, 2022
Multi-modal co-attention for drug-target interaction annotation and Its Application to SARS-CoV-2

CoaDTI Multi-modal co-attention for drug-target interaction annotation and Its Application to SARS-CoV-2 Abstract Environment The test was conducted i

Layne_Huang 7 Nov 14, 2022
Python wrappers to the C++ library SymEngine, a fast C++ symbolic manipulation library.

SymEngine Python Wrappers Python wrappers to the C++ library SymEngine, a fast C++ symbolic manipulation library. Installation Pip See License section

136 Dec 28, 2022
Simulation of the solar system using various nummerical methods

solar-system Simulation of the solar system using various nummerical methods Download the repo Make shure matplotlib, scipy etc. are installed execute

Caspar 7 Jul 15, 2022
MAGMA - a GPT-style multimodal model that can understand any combination of images and language

MAGMA -- Multimodal Augmentation of Generative Models through Adapter-based Finetuning Authors repo (alphabetical) Constantin (CoEich), Mayukh (Mayukh

Aleph Alpha GmbH 331 Jan 03, 2023
Code release for Local Light Field Fusion at SIGGRAPH 2019

Local Light Field Fusion Project | Video | Paper Tensorflow implementation for novel view synthesis from sparse input images. Local Light Field Fusion

1.1k Dec 27, 2022
A hyperparameter optimization framework

Optuna: A hyperparameter optimization framework Website | Docs | Install Guide | Tutorial Optuna is an automatic hyperparameter optimization software

7.4k Jan 04, 2023
The story of Chicken for Club Bing

Chicken Story tl;dr: The time when Microsoft banned my entire country for cheating at Club Bing. (A lot of the details are from memory so I've recreat

Eyal 142 May 16, 2022
BERTMap: A BERT-Based Ontology Alignment System

BERTMap: A BERT-based Ontology Alignment System Important Notices The relevant paper was accepted in AAAI-2022. Arxiv version is available at: https:/

KRR 36 Dec 24, 2022
QRec: A Python Framework for quick implementation of recommender systems (TensorFlow Based)

Introduction QRec is a Python framework for recommender systems (Supported by Python 3.7.4 and Tensorflow 1.14+) in which a number of influential and

Yu 1.4k Dec 30, 2022
Code for the SIGGRAPH 2022 paper "DeltaConv: Anisotropic Operators for Geometric Deep Learning on Point Clouds."

DeltaConv [Paper] [Project page] Code for the SIGGRAPH 2022 paper "DeltaConv: Anisotropic Operators for Geometric Deep Learning on Point Clouds" by Ru

98 Nov 26, 2022
git《Self-Attention Attribution: Interpreting Information Interactions Inside Transformer》(AAAI 2021) GitHub:

Self-Attention Attribution This repository contains the implementation for AAAI-2021 paper Self-Attention Attribution: Interpreting Information Intera

60 Dec 29, 2022
This repository contains tutorials for the py4DSTEM Python package

py4DSTEM Tutorials This repository contains tutorials for the py4DSTEM Python package. For more information about py4DSTEM, including installation ins

11 Dec 23, 2022