Refactor priors #329

dweindl · 2024-12-04T14:04:06Z

Introduces Prior and Distribution classes for handling PEtab-specific prior distributions, and (PEtab-version-invariant) univariate probability distributions. Supports sampling from them, and evaluating negative log-priors (#312). Later on, this can be extended to noise models for measurements and computing loglikelihoods.
This also adds a notebook demonstrating the various prior options which are a common source confusion.

Closes #311.

👀 notebook: https://petab--329.org.readthedocs.build/projects/libpetab-python/en/329/example/distributions.html

codecov-commenter · 2024-12-04T14:18:26Z

Codecov Report

Attention: Patch coverage is 82.63158% with 33 lines in your changes missing coverage. Please review.

Project coverage is 74.65%. Comparing base (d7f7e3a) to head (7ecc231).

Files with missing lines	Patch %	Lines
petab/v1/priors.py	84.09%	9 Missing and 5 partials ⚠️
petab/v1/distributions.py	88.50%	10 Missing ⚠️
petab/v1/sampling.py	25.00%	8 Missing and 1 partial ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##           develop     #329      +/-   ##
===========================================
+ Coverage    74.51%   74.65%   +0.13%     
===========================================
  Files           53       54       +1     
  Lines         5222     5381     +159     
  Branches       910      923      +13     
===========================================
+ Hits          3891     4017     +126     
- Misses         984     1012      +28     
- Partials       347      352       +5

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

dweindl · 2024-12-09T13:46:40Z

Requires some clarification related to normalization of the prior density. Related to PEtab-dev/PEtab#402.

EDIT: Needs clarification, but the current implementation matches what was done before.

dilpath

Looks good, thanks 👍

petab/v1/distributions.py

dilpath · 2024-12-09T21:53:13Z

petab/v1/distributions.py

+    def __init__(
+        self,
+        type_: str,
+        transformation: str,
+        parameters: tuple,
+        bounds: tuple = None,
+        parameter_scale: bool = False,
+    ):


Support supplying some rng: np.Generator = None and otherwise set self.rng = np.random.default_rng() in __init__?

Makes sense. I'll add that separately. For now, it's handled the same way as before, using the default numpy random state.

petab/v1/distributions.py

petab/v1/sampling.py

dilpath · 2024-12-09T22:10:45Z

tests/v1/test_distributions.py

+
+    # Kolmogorov-Smirnov test to check if the sample is drawn from the CDF
+    _, p = kstest(sample, cdf)
+    assert p > 0.05, (p, distribution)


Should this be higher than 0.05? e.g. p > 0.5 should also always work for 10000 samples, just guessing. Anyway, fine as is.

With p > 0.5 this will fail frequently. Even with p > 0.05 it fails often enough so that I set a seed.

That's surprising! I would have guessed that 10,000 samples from a "simple" distribution should be enough to say "these distributions look similar"... anyway fine 👍

Introduce `Distribution` classes for handling prior distributions and sampling from them. Later on this can be extended to noise models for measurements.

dilpath

Looks neat, thanks!

dilpath · 2024-12-10T23:39:05Z

petab/v1/distributions.py

+    def __init__(self, log: bool | float = False):
+        if log is True:
+            log = np.exp(1)
+        self._log = log


I find this more readable in _undo_log and __apply_log

Suggested change

self._log = log

self._log_base = log

dilpath · 2024-12-10T23:47:15Z

petab/v1/priors.py

+        match type_, transformation:
+            case (C.UNIFORM, _) | (C.PARAMETER_SCALE_UNIFORM, C.LIN):
+                self.distribution = Uniform(*parameters)
+            case (C.NORMAL, _) | (C.PARAMETER_SCALE_NORMAL, C.LIN):
+                self.distribution = Normal(*parameters)
+            case (C.LAPLACE, _) | (C.PARAMETER_SCALE_LAPLACE, C.LIN):
+                self.distribution = Laplace(*parameters)
+            case (C.PARAMETER_SCALE_UNIFORM, C.LOG):
+                self.distribution = Uniform(*parameters, log=True)
+            case (C.LOG_NORMAL, _) | (C.PARAMETER_SCALE_NORMAL, C.LOG):
+                self.distribution = Normal(*parameters, log=True)
+            case (C.LOG_LAPLACE, _) | (C.PARAMETER_SCALE_LAPLACE, C.LOG):
+                self.distribution = Laplace(*parameters, log=True)
+            case (C.PARAMETER_SCALE_UNIFORM, C.LOG10):
+                self.distribution = Uniform(*parameters, log=10)
+            case (C.PARAMETER_SCALE_NORMAL, C.LOG10):
+                self.distribution = Normal(*parameters, log=10)
+            case (C.PARAMETER_SCALE_LAPLACE, C.LOG10):
+                self.distribution = Laplace(*parameters, log=10)
+            case _:
+                raise ValueError(
+                    "Unsupported distribution type / transformation: "
+                    f"{type_} / {transformation}"
+                )


base 2? 😀

Once in PEtab 🙄

dilpath · 2024-12-10T23:50:48Z

tests/v1/test_distributions.py

+    # pdf -> cdf
+    def cdf(x):
+        return cumulative_trapezoid(distribution.pdf(x), x)


Makes sense to ensure that the samples match the PDF within PEtab. For the LIN transform cases, could you additionally compare to scipy.stats.(log)(uniform/normal/laplace)? I guess the LOG and LOG10 transforms will be dropped in v2 so less important to test I guess

dweindl force-pushed the 311_refactor_priors branch 5 times, most recently from 5313345 to b28e5c1 Compare December 4, 2024 22:38

dweindl marked this pull request as ready for review December 4, 2024 22:47

dweindl requested review from m-philipps and a team as code owners December 4, 2024 22:48

dweindl self-assigned this Dec 4, 2024

dweindl force-pushed the 311_refactor_priors branch 4 times, most recently from 721ba31 to d511639 Compare December 7, 2024 10:04

dweindl mentioned this pull request Dec 7, 2024

Implement proper truncation for prior distributions #335

Open

dweindl marked this pull request as draft December 8, 2024 11:25

dweindl force-pushed the 311_refactor_priors branch from d511639 to 9381d7c Compare December 9, 2024 13:42

dweindl mentioned this pull request Dec 9, 2024

Fix test_priors_to_measurements #336

Merged

dweindl force-pushed the 311_refactor_priors branch 4 times, most recently from a272255 to 86643ec Compare December 9, 2024 21:43

dweindl marked this pull request as ready for review December 9, 2024 21:44

dilpath approved these changes Dec 9, 2024

View reviewed changes

dweindl force-pushed the 311_refactor_priors branch 3 times, most recently from a6429c8 to 9013648 Compare December 10, 2024 08:25

Refactor priors

e25785f

Introduce `Distribution` classes for handling prior distributions and sampling from them. Later on this can be extended to noise models for measurements.

dweindl force-pushed the 311_refactor_priors branch from 9013648 to e25785f Compare December 10, 2024 08:26

once more

6420d67

dweindl force-pushed the 311_refactor_priors branch 2 times, most recently from 0b704ba to cc6a7e7 Compare December 10, 2024 17:30

dweindl requested a review from dilpath December 10, 2024 17:31

..

047dc46

dweindl force-pushed the 311_refactor_priors branch from cc6a7e7 to 047dc46 Compare December 10, 2024 17:41

dilpath approved these changes Dec 10, 2024

View reviewed changes

apply suggestions

7ecc231

dweindl merged commit 8456635 into PEtab-dev:develop Dec 11, 2024
7 checks passed

dweindl deleted the 311_refactor_priors branch December 11, 2024 08:49

dweindl mentioned this pull request Dec 11, 2024

Compute log-prior and unnormalized log-posterior in simulate_petab AMICI-dev/AMICI#2624

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Refactor priors #329

Refactor priors #329

dweindl commented Dec 4, 2024 •

edited

Loading

codecov-commenter commented Dec 4, 2024 •

edited

Loading

dweindl commented Dec 9, 2024 •

edited

Loading

dilpath left a comment

dilpath Dec 9, 2024

dweindl Dec 10, 2024

dilpath Dec 9, 2024

dweindl Dec 10, 2024

dilpath Dec 10, 2024

dilpath left a comment

dilpath Dec 10, 2024

dilpath Dec 10, 2024

dweindl Dec 11, 2024

dilpath Dec 10, 2024

Refactor priors #329

Refactor priors #329

Conversation

dweindl commented Dec 4, 2024 • edited Loading

codecov-commenter commented Dec 4, 2024 • edited Loading

Codecov Report

dweindl commented Dec 9, 2024 • edited Loading

dilpath left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

dilpath left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

dweindl commented Dec 4, 2024 •

edited

Loading

codecov-commenter commented Dec 4, 2024 •

edited

Loading

dweindl commented Dec 9, 2024 •

edited

Loading