Bayesian post-analysis using BUMPS DREAM by rozyczko · Pull Request #244 · easyscience/core

rozyczko · 2026-05-06T12:30:37Z

A new Bayesian MCMC sampling interface. This enables users to perform posterior sampling using the BUMPS DREAM sampler directly from both the MultiFitter and BUMPS minimizer classes.

Added a sample method to MultiFitter that flattens multi-dataset arrays, resolves parameter aliases, and delegates to the minimizer’s sampling method, supporting both 1D and vectorized multi-dimensional fits.
Implemented a sample method in the BUMPS minimizer (minimizer_bumps.py) that builds a BUMPS FitProblem and runs the DREAM sampler, returning posterior draws, parameter names, sampler state, and log-probabilities.

of bayesian

codecov · 2026-05-12T11:53:24Z

Codecov Report

❌ Patch coverage is 89.28571% with 9 lines in your changes missing coverage. Please review.
✅ Project coverage is 82.65%. Comparing base (d4a4c23) to head (e397973).

Files with missing lines	Patch %	Lines
.../easyscience/fitting/minimizers/minimizer_bumps.py	86.15%	9 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##           develop     #244      +/-   ##
===========================================
+ Coverage    82.55%   82.65%   +0.10%     
===========================================
  Files           62       62              
  Lines         4946     5023      +77     
===========================================
+ Hits          4083     4152      +69     
- Misses         863      871       +8

Flag	Coverage Δ
integration	`43.87% <70.23%> (+0.38%)`	⬆️
unittests	`82.14% <88.09%> (+0.86%)`	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
src/easyscience/fitting/fitter.py	`97.27% <100.00%> (+0.37%)`	⬆️
src/easyscience/fitting/multi_fitter.py	`98.55% <100.00%> (+1.40%)`	⬆️
.../easyscience/fitting/minimizers/minimizer_bumps.py	`95.95% <86.15%> (-3.52%)`	⬇️

rozyczko · 2026-05-19T09:16:05Z

I might want to extend this to multiprocessing, since BUMPS already supports this.
The implementation should speed things up quite noticeably.
(in a separate branch/PR)

henrikjacobsenfys

Very nice, found no issues except a few things in the notebook:

It's using ObjBase, which we're deprecating
Yo ucould add the hash to the pooch retrieve call

henrikjacobsenfys · 2026-05-29T08:18:25Z

Some of the notes don't get rendered right now

rozyczko · 2026-05-29T08:44:49Z

Some of the notes don't get rendered right now

hmmm
for me it displays the table nicely

henrikjacobsenfys · 2026-05-29T08:49:55Z

Some of the notes don't get rendered right now

hmmm for me it displays the table nicely

The tables are fine, it's the markdown that I'm unsure of. But could just be a vs code thing

rozyczko · 2026-05-29T08:59:28Z

Fixed both issues

damskii9992 · 2026-06-01T09:26:01Z

It's using ObjBase, which we're deprecating

Sadly we don't have the replacement yet. We need a new class for these kinds of easyscience standalone fits where you can just feed the class the parameters and the model function. To make it easy. Using ModelBase requires defining a class which is not something a user should be doing . . . .

So I think we should stick with using ObjBase (at least in tutorials) until we have created the replacement class.

damskii9992

I have only reviewed the notebook so far, but I feel like there is enough here to work on while I review the code.

damskii9992 · 2026-06-01T09:05:03Z

+    "\n",
+    "    return pooch.retrieve(\n",
+    "        url=f'https://public.esss.dk/groups/scipp/dmsc-summer-school/2025/{name}',\n",
+    "        known_hash=known_hash,\n",


We should probably have a .datasets folder in easyscience to make import of data-files for tutorials easy. A tutorial really shouldn't have cells like these . . .

You mean relying on dmsc-summer-school location or using pooch.retrieve?

I mean something like what I have in my documentation EasyImaging branch.

A datasets module with loader functions, which loads in files from its own pooch registry.

This, again, is a good refactoring, which needs to be done for all the tutorials, hence outside the scope of this PR. Maybe we can move it out to EasyUtils

#256

damskii9992 · 2026-06-01T10:30:02Z

+    "ax.plot(omega, mid, '-', color='C1', label='posterior median')\n",
+    "ax.set(xlabel='$\\\\omega$/meV', ylabel='$I(\\\\omega)$')\n",
+    "ax.legend()\n",
+    "plt.show()"


This should really be hidden away by a method in easyscience to make it a simple call like:

results.plot_model(confidence_intervals = [2.5, 97.5])

Or something like that in terms if the user-facing API.

As mentioned above - we don't want the core to provide us with plotting. Users choose what plotting mechanism they use with the data they get from fitting. Hence - no plopp/matplotlib/plotly dependence in the core itself.

damskii9992

Comments after reading the code and not just the user-facing API.

damskii9992 · 2026-06-01T12:19:18Z

+            pytest.skip('BUMPS is not installed')
+
+        result = f.sample(
+            x=[x2D], y=[y2D], weights=[weights], samples=100, burn=20, thin=2, vectorized=True


What is this vectorized keyword? And since we don't expose it in our API, why do we even test for it?

vectorized is indeed in mcmc_sample signature. It is quite important so I'll leave it as is

Can you then add a comment to the test to explain why/what? And should it be exposed in the API?

It is not just in sampling but in regular fitting as well. Nothing new.
But I added more comments explaining the usage cases.

damskii9992 · 2026-06-02T07:55:24Z

It's using ObjBase, which we're deprecating

Sadly we don't have the replacement yet. We need a new class for these kinds of easyscience standalone fits where you can just feed the class the parameters and the model function. To make it easy. Using ModelBase requires defining a class which is not something a user should be doing . . . .

So I think we should stick with using ObjBase (at least in tutorials) until we have created the replacement class.

Actually, come to think of it, I believe we can replace it with an EasyList for now. All the class needs to do is hold the parameters and have an get_all_fittable_parameters method implemented. I think that would be cleaner.

rozyczko · 2026-06-02T09:11:39Z

It's using ObjBase, which we're deprecating

Sadly we don't have the replacement yet. We need a new class for these kinds of easyscience standalone fits where you can just feed the class the parameters and the model function. To make it easy. Using ModelBase requires defining a class which is not something a user should be doing . . . .
So I think we should stick with using ObjBase (at least in tutorials) until we have created the replacement class.

Actually, come to think of it, I believe we can replace it with an EasyList for now. All the class needs to do is hold the parameters and have an get_all_fittable_parameters method implemented. I think that would be cleaner.

As discussed, I'll move back to ObjBase so all the tutorials can be modified to use EasyList at the same time, in a separate PR

#253

damskii9992

LGTM for now.

Integrate the squash-merged Bayesian DREAM feature from develop (PR #244) while retaining this branch's multiprocessing and pickling additions. Reconciliation notes: - Adopt develop's polymorphic `Fitter.mcmc_sample` (replaces the branch's separate `MultiFitter.sample`) and develop's API naming (`mcmc_sample`, return key `internal_bumps_object`) and input validation. - Extend `Fitter.mcmc_sample` with `chains`, `seed`, and `n_workers`, forwarding them to the BUMPS minimizer so multiprocessing works for both single- and multi-dataset fitters. - Keep the minimizer-level multiprocessing machinery: `BumpsPoolMapper`, worker helpers, cloudpickle-based problem serialization, the `chains` alias / `seed` handling, and `n_workers` wiring with mapper cleanup. - Merge the test suites: develop's mcmc_sample tests plus the branch's MP, pickling, alias, and seed tests (adapted to the mcmc_sample API). Make the n_workers subprocess test import the model by file path instead of via the shadowable `tests` namespace package. - Update the sampling benchmark tool to call `mcmc_sample`. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

initial implementation

c4c1b02

rozyczko added [scope] enhancement Adds/improves features (major.MINOR.patch) [priority] medium Normal/default priority labels May 6, 2026

rozyczko changed the title ~~initial implementation~~ Bayesian post-analysis using BUMPS DREAM May 6, 2026

rozyczko added 5 commits May 9, 2026 22:20

show max iteration, so users know how long to wait. Added cancellation

96be6d0

of bayesian

use absolute paths for wheel

181d43d

additional tests for bayesian

06cefc7

Merge branch 'develop' into bayesian

027552b

proper docstrings

efd7ff1

rozyczko added 4 commits May 12, 2026 17:11

we need legacy folder in the wheel

2913ca6

PR review issues addressed

8aa06c1

msg -> reason for exceptions

131f7c6

PR fixes

916b1d3

rozyczko marked this pull request as ready for review May 18, 2026 09:46

Merge branch 'develop' into bayesian

f072fd6

rozyczko mentioned this pull request May 20, 2026

Add sampling method for multiffiter/fitter #246

Open

rozyczko added 2 commits May 28, 2026 08:27

added bayesian notebook

e7f345a

updated/enhanced bayesian notebook

1088a90

henrikjacobsenfys approved these changes May 29, 2026

View reviewed changes

PR review of the notebook

db00f9f

damskii9992 reviewed Jun 1, 2026

View reviewed changes

rozyczko mentioned this pull request Jun 2, 2026

Move all the examples and tutorials to use EasyList #253

Open

rozyczko added 2 commits June 2, 2026 11:26

PR review comments - part 1

e912f5c

more PR review comments addressed

9395b8f

rozyczko requested a review from damskii9992 June 2, 2026 13:11

fixed bayesian failing...

e397973

damskii9992 approved these changes Jun 3, 2026

View reviewed changes

rozyczko merged commit aadbd48 into develop Jun 3, 2026
34 checks passed

Conversation

rozyczko commented May 6, 2026

Uh oh!

codecov Bot commented May 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

rozyczko commented May 19, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

henrikjacobsenfys left a comment

Choose a reason for hiding this comment

Uh oh!

henrikjacobsenfys commented May 29, 2026

Uh oh!

rozyczko commented May 29, 2026

Uh oh!

henrikjacobsenfys commented May 29, 2026

Uh oh!

rozyczko commented May 29, 2026

Uh oh!

damskii9992 commented Jun 1, 2026

Uh oh!

damskii9992 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

rozyczko Jun 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

damskii9992 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

damskii9992 commented Jun 2, 2026

Uh oh!

rozyczko commented Jun 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

damskii9992 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

codecov Bot commented May 12, 2026 •

edited

Loading

rozyczko commented May 19, 2026 •

edited

Loading

rozyczko Jun 2, 2026 •

edited

Loading

rozyczko commented Jun 2, 2026 •

edited

Loading