ubc-stat-ml / blangsdk

Blang's software development kit

Summary

Prospective/current users: please vist the project web page for more information as well as our JSS paper.

Docker image: link to documentation

New feature: Blang can now run on 1000s of machines using MPI via the Pigeons-Blang bridge

Blang developers: in addition to the above resources, see also the documentation repository for more information.

This is one of the repositories hosting Blang's code. This one contains the Blang's SDK (Software Development Kit), including:

• Basic datatypes suitable for sampling.
• Infrastructure to create new data types and distributions.
• Inference algorithms for such datatypes, such as Adaptive Non-Reversible Parallel Tempering and Sequential Change of Measure.
• Standard probability distributions.
• MCMC testing infrastructure.
• Runtime to perform static analysis to infer the factor graph and its sparsity patterns.
• Automated post-processing facilities (MCMC diagnostic, trace/density/pmf/summaries generation, etc).

See this readme for a roadmap of the other key repositories (language infrastructure, examples, supporting libraries, etc)

Citing Blang: if you find Blang useful for your work, consider citing our JSS paper:

Alexandre Bouchard-Côté, Kevin Chern, Davor Cubranic, Sahand Hosseini, Justin Hume, Matteo Lepur, Zihui Ouyang, Giorgio Sgarbi (2022)
Journal of Statistical Software 103:1–98

blangsdk's Issues

Further improvements on data frames

• Inverted control with plates
• Make Table interface, concrete types are ScalarTable and MatrixTable
  • ScalarTable configured by matrix.inPlates(..) only
  • MatrixTable configured by matrix.inPlates(..); matrix.rowsIndexedBy(...); matrix.colsIndexedBy(...)
  • In both cases, plate indices are modelled by sets (no duplicates)
• Keep the same idea in ScalarTable
• Sub-frames

Implement "primitive" types

• Real
• Int
• Bool

Associated:

• basic sampler
• command line parser

Add checks that all variables are covered by samplers

Phylo example

Need to add zero and one-string constructors to built-in datatypes

Simplex type, Dirichlet distribution and associated sampler

Demonstrate at same time that the type engineering architecture works

Reorganize examples

• Remove blang.examples from SDK
• Create project template
• Create some script to upgrade everything or via snapshots
• Write examples via an instantiation of the template

Add support for distributed SMC/PT

Move runtime from bayonet to blangSDK

This is required since these depend on interfaces such as Model that are needed in blangDSL (and it's non-trivial because of plug-in architecture and seems suboptimal to have blangDSL import bayonet).

Note: important to also erase from bayonet to avoid having two versions of same thing.

Clustering data structure, DP prior and split-merge sampler

Add static time conditionals?

Adaptation

Need to at least set univariate scale.

Sampler exclusions

Modify parsers

Consider having the parsers for matrices reading formats like [a b]

Check cases where a complex data structure is partially observed

E.g. are cases like Simplex correctly handled?

Basic HMM model and associated sampler

Demonstrate at same time reuse of models.

Move syntax tests from DSL to SDK

Already deleted from DSL as they were not catching basic syntax errors.

Some basic performance benchmarks

Standalone runner should handle file names with non standard characters gracefully

E.g., dashes

Exported maven artefacts too big

Currently almost 100MB per version. Will break artefact git repo soon if not fixed

Use xtend's range operators

E.g., in src/main/java/blang/examples/MixtureModel.bl, instead of StaticUtils::range(observations.size()), we should be able to use 0 ..< observations.size(). (Xtend reference)

Improve slice sampler

Use doubling instead of linear expansion procedure, re-read Neal's paper and re-write the implementation.

Upgrade dependencies

xtend, xtend-gradle, xtext, eclipse, maven dependencies

Upgrade inits and improve command line option reporting

Multivariate normal distribution, associated sampler

Table's should not be @Immutable

Basic HMM model and associated sampler - DUPLICATED

Demonstrate at same time reuse of models.

Simplify Factor class hierarchy

Get rid of LogScaleFactor, move its method to Factor. Generalize when (if) the need arises.

Improve publishing architecture - current artifact/git hack not scalable

Problems with current solution:

• Limitations of github pages: https://help.github.com/articles/what-is-github-pages/
  • max of 1GB
  • max of 10 builds per day
• Limitations of git: push gets slower

Add model integrity checks

• DAG property [implemented, still need to hook up?]
• all declared variables used [done? check]
• ..

Fix registration of names in GraphAnalysis

The annotation @DeboxedName should be used instead of the raw field name. Wait until variable printing architecture is setup.

Add support for enums

Do not do via java enums! Because the values cannot be read from command line then. Would also make a mess everywhere. Instead, use some kind of FiniteSupport

Need annealing to take care of initialization; and parallel tempering at same time for perf?

Issues:

• how to do the factor transformation
• transform them into what (what exponents?)
• how to mask bad energies from processors?

Circular dependency between an Xtend and Xtext not supported in gradle

E.g. in 187d1ad

in xtext type MarkovModel:

@Sampler(BaumWelch)

and in BaumWelch.xtend:

MarkovModel model

NB: works fine in eclipse, but breaks in gradle plugin as of latest version 1.0.12

Basic set of discrete distributions, associated samplers

Integer valued:

• Multinomial
• Poisson

Consider a new type of support factor with an explicit finite list. Allows better samplers.

Integrate xlinear

Most of the work will be with integration of graph introspection and views.

Light weight data frame structure needed for plated random variables

Working on Plated<T>. More details soon.

Add checks that static time constraints are constant

Add inference by enumeration?

More advanced tempering features

• See md document in /notes/

Add back slice sampler and/or non-reversible stuff

Reporting acceptance rates, etc

Improve data frame

• Most important: make it so that loops can be combined,
• Conversion into lists/matrices

Improve performance of cloning

• Some immutable stuff to mark for shallow copy, e.g. arrays of integers for sampler cache
• Datasets
• Plates, etc [mark as immutable as well]
• Parallelize!

Implement parallel tempering

!! Create compute-aware annealing methods

E.g. for trees or exchangeable datasets, use only subset of sites in the first iterations

Basic set of real-valued distributions and better sampler (say slice at least)

• Normal
• Gamma
• Beta
• ...

Command-line blang has a 5--10 sec latency

@justinHume : I'll look into that

Fix issues encountered while testing MixtureModel

• Bug: posteriors clearly incorrect
• Performance: seems very slow at 1000 datapoint, encouragingly, get much more than 2x faster at 500 datapoints, so some factor dependencies must be wrong
• Marking fields with transient seems to cause undesirable side effects, need better way to signal graph inference to skip some [document adding transient to K value changes behavior]
• Style: problematic that loops cannot be shared [split to separate issue]
• Style: autodeboxing is needed [see separate issue]

Fix printing

• Make it work on SMC samplers
• Use hierarchical dir? Field names? ExperimentResults?