StanSample.jl

WIP: Wrapper package for the sample method in Stan's cmdstan executable.
Author StanJulia
Popularity
13 Stars
Updated Last
11 Months Ago
Started In
June 2019

StanSample v7

Project Status Build Status

Purpose

StanSample.jl wraps cmdstan's sample method to generate draws from a Stan Language Program. It is the primary workhorse in the StanJulia ecosystem.

StanSample.jl v7 supports InferenceObjects.jl as a package extension. Use inferencedata(model) to create an InferenceData object. See also note 1 below. An example Pluto notebook can be found here

Notes

  1. Use of both InferenceObjects.jl and the read_samples() output_format options :dimarray and :dimarrays (based on DimensionalData.jl) creates a conflict. Hence these output_format options are no longer included. See the example Pluto notebook test_dimarray.jlin StanExampleNotebooks.jl for an example how to still use that option. At some point in time InferenceObjects.jl might provide an alternative way to create a stacked DataFrame and/or DimensionalData object.

  2. I've removed BridgeStan.jl from StanSample.jl. Two example Pluto notebooks, test_bridgestan.jl and bridgestan_stansample_example.jl in StanExampleNotebooks.jl demonstrate how BridgeStan can be used.

Prerequisites

You need a working installation of Stan's cmdstan, the path of which you should specify in either CMDSTAN or JULIA_CMDSTAN_HOME, e.g. in your ~/.julia/config/startup.jl include a line like:

# CmdStan setup
ENV["CMDSTAN"] =
     expanduser("~/.../cmdstan/") # replace with your path

Or you can define and export CMDSTAN in your .profile, .bashrc, .zshrc, etc.

For more details see this file.

See the example/bernoulli.jl for a basic example. Many more examples and test scripts are available in this package and also in Stan.jl.

Multi-threading and multi-chaining behavior.

From StanSample.jl v6 onwards 2 mechanisms for in paralel drawing samples for chains are supported, i.e. on C++ level (using threads) and on Julia level (by spawning a Julia process for each chain).

The use_cpp_chains keyword argument in the call to stan_sample() determines if chains are executed on C++ level or on Julia level. By default, use_cpp_chains = false.

From cmdstan-2.28.0 onwards it is possible to use C++ threads to run multiple chains by setting use_cpp_chains=true in the call to stan_sample():

rc = stan_sample(_your_model_; use_cpp_chains=true, [ data | init | ...])

To enable multithreading in cmdstan specify this before the build process of cmdstan, i.e. before running make -j9 build. I typically create a path_to_my_cmdstan_directory/make/local file containing STAN_THREADS=true. You can see an example in .github/CI.yml script.

By default in either case num_chains=4. See ??stan_sample for all keyword arguments. Internally, num_chains will be copied to either num_cpp_chains or num_julia_chains.

Currently I do not suggest to use both C++ and Julia level chains. Based on the value of use_cpp_chains (true or false) the stan_sample() method will set either num_cpp_chains=num_chains; num_julia_chains=1 or num_julia_chains=num_chains;num_cpp_chain=1.

This default behavior can be disabled by setting the postional check_num_chains argument in the call to stan_sample() to false.

Threads on C++ level can be used in multiple ways, e.g. to run separate chains and to speed up certain operations. By default StanSample.jl's SampleModel sets the C++ num_threads to 4.

See the (updated for cmdstan-2.29.0) RedCardsStudy example graphs in Stan.jl and here for more details, in particular with respect to just enabling threads and including TBB or not on Intel, and also some indications of the performance on an Apple's M1/ARM processor running native (not using Rosetta and without Intel's TBB).

In some cases I have seen performance advantages using both Julia threads and C++ threads but too many combined threads certainly doesn't help. Note that if you only want 1000 draws (using 1000 warmup samples for tuning), multiple chains (C++ or Julia) do not help.

Installation

This package is registered. It can be installed with:

pkg> add StanSample.jl

Usage

Use this package like this:

using StanSample

See the docstrings (in particular ??StanSample) for more help.

Versions

Version 7.1-4.0

  1. Switch to cmdstan.2.32.0 for testing
  2. Removed BridgeStan extension

Version 7.0.1

  1. Updated column types for sample_stats (NamedTuples and DataFrames)

Version 7.0.0

  1. InferenceObjects.jl support.
  2. Conditional support for BridgeStan.
  3. Reduced support for :dimarray and :dimarrays option in read_samples().

Version 6.13.8

  1. Support for InferenceObjects v0.3.
  2. Many tmp directories created during testing have been removed from the repo.
  3. Support for BridgeStan v1.0 has been dropped.

Version 6.13.7

  1. Moved InferenceObjects behind Requires
  2. Method inferencedata() is using inferencedata3() currently

Version 6.13.6

  1. Added inferencedata3()
  2. Added option to enable logging in the terminal (thanks to @FelixNoessler)

Version 6.13.0 - 6.13.5

  1. Many more (minor and a bit more) updates to inferencedata()
  2. Updates to BridgeStan (more to be expected soon)
  3. Fix for chain numbering when using CPP threads (thanks to @apinter)
  4. Switched to use cmdstan-2.32.0 for testing
  5. Updates to Examples_Notebooks (in particular now using both inferencedata() and inferencedata2())
  6. Dropped support for read_samples(m, :dimarray) as this conflicted with InferenceData

Version 6.12.0

  1. Added experimental version of inferencedata(). See example in ./test/test_inferencedata.jl
  2. Added InferenceObjects.jl as a dependency
  3. Dropped MonteCarloMeasurements.jl as a dependency (still supported using Requires)
  4. Dropped MCMCChains.jl as a dependency (still supported using Requires)
  5. Dropped AxisKeys.jl as a dependency

Version 6.11.5

  1. Add sig_figs field to SampleModel (thanks to Andrew Radcliffe).

This change enables the user to control the number of significant digits which are preserved in the output. sig_figs=6 is the default cmdstan option, which is what StanSample has been defaulting to.

Typically, a user should prefer to generate outputs with sig_figs=18 so that the f64's are uniquely identified. It might be wise to make such a recommendation in the documentation, but I suppose that casual users would complain about the correspondingly increased .csv sizes (and subsequent read times).

Version 6.11.4

  1. Dropped conversion to Symbols in read_csv_files() if internals are requested (include_internals=true)
  2. Added InferenceObjects as a dependency.

This is part of the work with Set Haxen to enable working with InferenceData objects in a future release (probably v6.12).

Version 6.11.1

  1. Fix bridge_path in SampleModel.

Version 6.11.0

  1. Support for BridgeStan as a dependency of StanSample.jl (Thanks to Seth Axen)

Version 6.10.0

  1. Support for the updated version of BridgeStan.

Version 6.9.3

  1. A much better test has been added for multidimensional input arrays thanks to Andy Pohl (test/test_JSON).

Version 6.9.2

  1. More general handling of Array input data to cmdstan if the Array has more than 2 dimensions.

Version 6.9.2

  1. Experimental support for BridgeStan.

Version 6.9.0-1

  1. For chains read in as either a :dataframe or a :nesteddataframe the function matrix(...) has been replaced by array(...). Depending on the the eltype of the requested column, array will return a Vector, a Matrix or an Array with 3 dimensions.
  2. The function describe() has been added which returns a df with results based on Stan's stansummary executable.
  3. A new method has been added to DataFrames.getindex to extract cells in stansummary DataFrame, e.g. ss1_1[:a, :ess].

Version 6.8.0 (nesteddataframe is experimental!)

  1. Added :nesteddataframe option to read_samples(). Maybe useful if cmdstan returns vectors or matrices.
  2. Extended the matrix() function to matrix(df, Symbol).

Version 6.7.0

  1. Drops support for creating R files.
  2. Requires StanBase 4.7.0

Version 6.4.0

  1. Introduced available_chains("your model")
  2. Updated Redcardsstudy results for cmdstan-2.29.0

Version 6.3.0-1

  1. Switch to cmdstan-2.29.0 testing.

Version 6.2.1

  1. Better handling of .csv chain retrieval in read_csv_files.

Version 6.2.0

  1. Revert back to by default use Julia level chains.

Version 6.1.1-2

  1. Documentation improvements.

version 6.1.0

  1. Modified (simplified?) use of num_chains to define either number of chains on C++ or Julia level based on use_cpp_chains keyword argument to stan_sample().

Version 6.0.0

  1. Switch to C++ threads by default.
  2. Use JSON3.jl for data.json and init.json as replacement for data.r and init.r files.
  3. The function read_generated_quantities() has been dropped.
  4. The function stan_generate_quantites() now returns a DataFrame.

Version 5.4 - 5.6

  1. Full usage of num_threads and num_cpp_threads

Version 5.3.1 & 5.3.2

  1. Drop the use of the STAN_NUM_THREADS environment variable in favor of the keyword num_threads in stan_sample(). Default value is 4.

Version 5.3

  1. Enable local multithreading. Local as cmdstan needs to be built with STAN_THREADS=true (see make/local examples).

Version 5.2

  1. Switch use CMDSTAN environment variable

version 5.1

  1. Testing with conda based install (Windows, but also other platforms)

Versions 5.0

  1. Docs updates.
  2. Fix for DimensionalData v0.19.1 (@dim no longer exported)
  3. Added DataFrame parameter blocking option.

Version 5.0.0

  1. Keyword based SampleModel and stan_sample().
  2. Dropped dependency on StanBase.
  3. Needs cmdstan 2.28.1 (for num_threads).
  4. tmpdir now positional argument in SampleZModel.
  5. Refactor src dir (add common subdir).
  6. stan_sample() is now an alias for stan_run().

Version 4.3.0

  1. Added keywords seed and n_chains to stan_sample().
  2. SampleModel no longer uses shared fields (prep work for v5).

version 4.2.0

  1. Minor updates
  2. Added test for MCMCChains

Version 4.1.0

  1. The addition of :dimarray and :dimarrays output_format (see ?read_samples).
  2. No longer re-exporting many previously exported packages.
  3. The use of Requires.jl to enable most output_format options.
  4. All example scripts have been moved to Stan.jl (because of item 3).

Version 4.0.0 (BREAKING RELEASE!)

  1. Make KeyedArray chains the read_samples() default output.
  2. Drop the output_format kwarg, e.g.: read_samples(model, :dataframe).
  3. Default output format is KeyedArray chains, i.e.: chns = read_samples(model).

Version 3.1.0

  1. Introduction of Tables.jl interface as an output_format option (:table).
  2. Overloading Tables.matrix to group a variable in Stan's output file as a matrix.
  3. Re-used code in read_csv_files() for generated_quantities.
  4. The read_samples() method now consistently applies keyword arguments start and chains.
  5. The table for each chain output_format is :tables.

Version 3.0.1

  1. Thanks to the help of John Wright (@jwright11) all StanJulia packages have been tested on Windows. Most functionality work, with one exception. Stansummary.exe fails on Windows if warmup samples have been saved.

Version 3.0.0

  1. By default read_samples(model) will return a NamedTuple with all chains appended.
  2. output_format=:namedtuples will provide a NamedTuple with separate chains.

Version 2.2.5

  1. Thanks to @yiyuezhuo, a function extract has been added to simplify grouping variables into a NamedTuple.
  2. read_sample() output_format argument has been extended with an option to request conversion to a NamedTuple.

Version 2.2.4

  1. Dropped the use of pmap in StanBase