Repository navigation
Conversation
nomadbl
requested review from
Laasya-73,
alegresor,
fjhickernell and
sou-cheng-choi
October 21, 2026 16:38
![github-code-quality[bot]](https://avatars.githubusercontent.com/u/9919?s=60&v=4){kind=small}
| f"'{_missing_module}'. Install the pymc >= 6.0.1" | ||
| ) | ||
|
|
||
| class PyMCWrapper(AbstractTrueMeasure): |
Comment on lines
+417
to
+466
| # def _detect_issues(icdf_f, dims, split_indices, n_grid=3): | ||
| # issues = [] | ||
|
|
||
| # eps = np.finfo(float).eps | ||
| # us = np.meshgrid(*[np.linspace(eps, 1.0 - eps, n_grid) for _ in range(dims)], indexing='ij') | ||
| # us = [u.ravel() for u in us] | ||
| # us = np.stack(us, axis=-1) | ||
| # us = np.split(us, split_indices, axis=-1) | ||
| # us = [_x.ravel() for _x in us] | ||
| # if np.prod(us[-1].shape) == 0: | ||
| # us = us[:-1] | ||
| # try: | ||
| # x = icdf_f(*us) | ||
| # except Exception as e: | ||
| # return [f"icdf raised {type(e).__name__}: {e}"] | ||
|
|
||
| # if np.any(~np.isfinite(x)): | ||
| # issues.append("ppf() returned non-finite values") | ||
|
|
||
| # # TODO: can we generalize this for dims>1? | ||
| # # dx = np.diff(x) | ||
| # # if np.any(dx < -1e-12): | ||
| # # issues.append("ppf() is not nondecreasing (looks non-monotone)") | ||
|
|
||
| # # Light-touch pdf/logpdf checks on quantiles (works for bounded/unbounded support) | ||
| # # uq = np.linspace(0.01, 0.99, 9) | ||
| # # try: | ||
| # # xq = np.asarray(icdf_f(uq), dtype=float) | ||
| # # except Exception: | ||
| # # xq = None | ||
|
|
||
| # # if xq is not None and hasattr(dist, "pdf"): | ||
| # # try: | ||
| # # p = np.asarray(dist.pdf(xq), dtype=float) | ||
| # # if np.any(~np.isfinite(p)): | ||
| # # issues.append("pdf() returned non-finite values on quantiles") | ||
| # # if np.any(p < -1e-12): | ||
| # # issues.append("pdf() returned negative values on quantiles") | ||
| # # except Exception as e: | ||
| # # issues.append(f"pdf() raised {type(e).__name__}: {e}") | ||
|
|
||
| # # if xq is not None and hasattr(dist, "logpdf"): | ||
| # # try: | ||
| # # lp = np.asarray(dist.logpdf(xq), dtype=float) | ||
| # # if np.any(~np.isfinite(lp)): | ||
| # # issues.append("logpdf() returned non-finite values on quantiles") | ||
| # # except Exception as e: | ||
| # # issues.append(f"logpdf() raised {type(e).__name__}: {e}") | ||
|
|
||
| # return issues |
Comment on lines
+607
to
+626
| # try: | ||
| # # do some sampling and check no errors | ||
| # bs = 10 | ||
| # inputs = np.ones((bs, self.d))*0.5 | ||
| # # split to right array lengths and remove trailing empty array if present | ||
| # if self.d > 1: | ||
| # inputs_list = np.split(inputs, self.split_indices, axis=-1) | ||
| # else: | ||
| # inputs_list = [inputs.ravel()] | ||
| # inputs_list = [_x.ravel() for _x in inputs_list] | ||
| # if np.prod(inputs[-1].shape) == 0: | ||
| # inputs_list = inputs_list[:-1] | ||
| # # reshape to correct shapes | ||
| # print(self.shapes) | ||
| # print(inputs_list) | ||
| # inputs_list = [np.reshape(_x, bs, *shape) for _x, shape in zip(inputs_list, self.shapes)] | ||
| # _ = self.icdf(*inputs_list) | ||
| # except: | ||
| # raise DimensionError( | ||
| # f"sampler dimension {sampler.d} must match the model joint samples dimension {self.dimension}." |
Comment on lines
+630
to
+635
| # if issues: | ||
| # warnings.warn( | ||
| # "PyMCWrapper received a PyMC Model " | ||
| # "and it failed sanity checks:\n - " + "\n - ".join(issues), | ||
| # UserWarning, | ||
| # stacklevel=2, |
|
|
||
| import numpy as np | ||
| import warnings | ||
| from pytensor.tensor.utils import get_static_shape_from_size_variables |
| self.models = [] | ||
| # define models for tests | ||
| with pymc.Model() as normal_model: | ||
| x = pymc.Normal("x", mu=0.0, sigma=1.0) |
| nu = 5.0 | ||
| with pymc.Model() as t_model: | ||
| s = pymc.ChiSquared("s", nu=nu) | ||
| x = pymc.Normal("x", mu=0.0, sigma=pymc.math.sqrt(nu/s)) |
|
|
||
| with pymc.Model() as observed_t_model: | ||
| s = pymc.ChiSquared("s", nu=nu) | ||
| x = pymc.Normal("x", mu=0.0, sigma=pymc.math.sqrt(nu/s), observed=np.ones(3)) |
Comment on lines
+8
to
+11
| from qmcpy import ( | ||
| PyMCWrapper, | ||
| DigitalNetB2, | ||
| ) |
Member
|
Kindly review PR#642 to your branch. It is an attempt to fix the unit test failures. Hope it helps as discussed in Slack with @alegresor . |
sou-cheng-choi
requested changes
Oct 22, 2026
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Evaluate support for composable hierarchical TrueMeasure models
This PR adds a
TrueMeasuresub-class that allows sampling from distributions constructed by PyMC.This substantially extends the class of distributions which can be sampled by QMCPy "out of the box", and potentially extends application areas (which is the author's intent). The main limitation from the PyMC side is some missing implementations of
icdfwhich can be (and is progressively) patched.demos/pymc_model.ipynbfor usage examples, limitations and workarounds for them.pymcandpytensorneeded for this feature.demos/pymc_model.ipynbdownloads data fromhttps://github.com/stan-dev/posteriordbto run one of the examples. This data is not commited to the repository.AI Assistance
Checklist
Implementation Overview
PyMC was chosen specifically as a widely used, fully featured, python based, framework. This was preferred over the initial proof-of-concept implementation.
PyMC is built on top of PyTensor. PyTensor constructs computational graphs which can be "compiled" (actually, linked) to a program. The distinguishing feature of PyTensor is the ability to transform graphs, so several different computations can be implicitly specified by a single graph that gets transformed in different ways.
PyMC graphs are composed of a combination of two kinds of nodes: random variables like
normal_rv, and deterministic operations likept.exp,pt.logetc.PyMC Models are (a collection of) these graphs, stored in the model's
basic_RVs,unobserved_RVsand a few others. PyMC contains utilities to transform the random variable nodes into the corresponding computations for the log probability, which is then used in the implementation of algorithms.This PR builds on PyMC's utilities, replacing the random variable nodes with
icdfnodes to transform the original graph into a graph performing inverse-transform-sampling. The code in the class itself compiles this transform by calling these utilities. Most of the code in the class itself is glue code transforming arrays to/from QMCPy expected format and the format ingested by the compiled transform function.Related utilities to add - separate PR or here
Sometimes models get really complicated, with a large number of variables of different shapes etc.
PyMC deals with this by defaulting outputs to a dictionary format
{"var_name": ndarray, ...}. QMCPy seems to require/assume flat vectors as the "currency", thus I defaultedPyMCWrapperto output such flat vectors and the user is expected to know how to reconstruct the variables.A useful utility to accompany here would be a "converter" that reconstructs the dictionary format from the flat format, and the converter would be automatically deduced from the
Modelpassed in. I assumed this was appropriate as a separate PR but this can be added here without much effort.