Parallel uniform sampling + New sampler interface

[segasai]

This is a preliminary version of patch that enables parallel uniform sampling.
The idea is to get away from this propose/evolve for uniform distributions where evolve doesn't really do anything other than evaluate logl on a single pt, but instead make propose a no-op for uniform, and evolve() actually do the sampling inside the boundary till a satisfactory pt is found.

While implementing this I realized that after some changes all the different samplers in nestedsamplers.py may not need to be there as they do something very similar (but that's tbd).

ATM some tests fail, but I think that's fixable.

Also some class members had to be renamed. I.e. .bound attribute is really misleading in the classes as it was storing the list of all the historic bounds, so I renamed that to bound_list.

[coveralls]

Pull Request Test Coverage Report for Build 7952108689

Details

• -12 of 225 (94.67%) changed or added relevant lines in 6 files are covered.
• 1 unchanged line in 1 file lost coverage.
• Overall coverage increased (+0.2%) to 91.863%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
py/dynesty/sampling.py	28	30	93.33%
py/dynesty/bounding.py	30	34	88.24%
py/dynesty/sampler.py	136	142	95.77%

Files with Coverage Reduction	New Missed Lines	%
py/dynesty/bounding.py	1	92.24%

Totals
Change from base Build 7180723466:	0.2%
Covered Lines:	4053
Relevant Lines:	4412

---

💛 - Coveralls

[joshspeagle]

!!!! This is an excellent proposed change! Very interested to see how the other internal refactoring might also work.

[segasai]

Thanks @joshspeagle
Just out of curiosity i've actually refactored nestedsamplers. It was surprisingly easy to get rid of those classes there which removed few hundred lines (of codes + comments).
I am not 100%, but 95% certain that's a good change just because it's easier to understand it now (and there was so little different between those samplers)

[segasai]

The demonstration of this patch. The following uniform sampling of Eggbox with single bound
takes 36 minutes in the current dynesty :
9528it [36:29, 4.35it/s, +1000 | bound: 8600 | nc: 1 | ncall: 12936194 | eff(%): 0.081 | loglstar: -inf < 243.000 < inf | logz: 235.824 +/- 0.079 | dlogz: 0.000 > 0.100]
With the patch it takes 3 minutes.

9490it [03:16, 48.25it/s, +1000 | bound: 999 | nc: 1 | ncall: 12943702 | eff(%): 0.081 | loglstar: -inf < 243.000 < inf | logz: 235.863 +/- 0.078 | dlogz: 0.000 > 0.100]

And if I bump the number of threads to 36 , it takes 50 seconds. I.e. basically it scales properly, while previously uniform sampler basically didn't really scale with multiple threads.

import numpy as np
import dynesty
import multiprocessing as mp

nlive = 1000


def loglike_egg(x):
    logl = ((2 + np.cos(x[0] / 2) * np.cos(x[1] / 2))**5)
    return logl


def prior_transform_egg(x):
    return x * 10 * np.pi


LOGZ_TRUTH = 235.856


def test_bounds():
    bound, sample = 'single', 'unif'
    # stress test various boundaries
    ndim = 2
    rstate = np.random.default_rng(444)
    with mp.Pool(4) as poo:
        sampler = dynesty.NestedSampler(loglike_egg,
                                        prior_transform_egg,
                                        ndim,
                                        nlive=nlive,
                                        bound=bound,
                                        sample=sample,
                                        rstate=rstate,
                                        pool=poo,
                                        queue_size=4)
        sampler.run_nested(dlogz=0.1, print_progress=True)
        assert (abs(LOGZ_TRUTH - sampler.results.logz[-1])
                < 5. * sampler.results.logzerr[-1])


if __name__ == '__main__':
    test_bounds()

[segasai]

It would be great to have some review of this patch @joshspeagle if you have time, before the patch gets bigger.

My current thinking is the patch is mostly positive with a few negatives listed below:

• Previously I had a test of custom sampler like that _SAMPLERS["custom"] = MultiEllipsoidSampler. I had to disable this, but I don't actually know if anyone in the wild is using this functionality.
• The new logic in propose_live has too many ifs() for my liking because it has to deal with combinations of different bounds/samplers. I think this needs to be cleaned up.
• kwargs is used to communicate the bound and a few other variables to sample_bound_unif. Maybe that needs to be changed and we need to expand SampleArgument

  dynesty/py/dynesty/sampling.py

  Line 24 in 3f10025

  SamplerArgument = namedtuple('SamplerArgument', [

  rather than abuse kwargs
• Renaming of a few attributes (there were .method/.sampling attributes, now there is only .sampling). The .bound is renamed to .bound_list.
• Some interface change to bounding classes, so they can be used interchangeably. Maybe having an BoundingInterface class is a good idea...
• I don't have an actual example, but I could imagine the case where the new implementation is slower. Maybe the case of very high acceptance rate, very fast logl function, and very large bound, in this case the overhead from pickling the bound and sending it to the pool worker may be significant. Probably this is not a big worry, since if acceptance fraction is high, things are easy no matter what.

Among things I'd like to do, but I don't want to do in this patch: I think with this patch it's becoming easier to do #391. We already now can pack update_slice or update_rwalk in SliceSampler RWalkSampler objects and make sure that things like history

dynesty/py/dynesty/nestedsamplers.py

Line 145 in 3f10025

self.slice_history = {'ncontract': 0, 'nexpand': 0}

is stored inside those objects. But I'd prefer to do that separately, as this requires more thought.

[Copilot]

Pull Request Overview

This PR introduces a new sampler interface for parallel uniform sampling while deprecating several legacy components. Key changes include replacing function‐based samplers with object‐instantiated versions, renaming several internal attributes (e.g. “bound” → “bound_list” and “M” → “mapper”), and removing or refactoring tests for deprecated custom update functionality.

Reviewed Changes

Copilot reviewed 15 out of 17 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
tests/test_sampling.py	Updated sampler references to use new object‐based sampler API.
tests/test_misc.py	Removed tests for custom sampler registration and update logic.
tests/test_gau.py	Adjusted formatting and removed gradient–related tests.
tests/test_ellipsoid.py	Updated ellipsoid Monte Carlo calls to use new “ctrs” attribute.
py/dynesty/utils.py	Renamed “M” attribute to “mapper” in restore_sampler method.
py/dynesty/plotting.py	Updated bound sampling calls to use “ctrs” instead of extra args.
py/dynesty/dynesty.py	Refactored sampler initialization and internal API (e.g. removed gradient argument in auto–selection, updated citations, renamed ncall).
py/dynesty/dynamicsampler.py	Refactored dynamic sampler internals including renaming bound→bound_list and updating pool mapping usage.
docs/source/conf.py	Minor comment update.
.readthedocs.yaml & .github/workflows/test.yml	Updated build environment and python versions.

Files not reviewed (2)

• docs/source/api.rst: Language not supported
• docs/source/references.rst: Language not supported

[mj-will]

Hi @segasai, this looks like a great update to dynesty!

The changes here have potential to break some codes, so this will be either 2.2 version, or 3.0. (

I just wanted to chime in on this as a downstream user; I'd strongly favour a 3.0 version if possible since it will make the maintenance on our side a lot easier and cleaner.

Thanks for continued work on dynesty.