Upgrade econml to 0.16.0 and dowhy to 0.14 (numpy 2, Python 3.10–3.12)#338
Merged
Conversation
Signed-off-by: Egor Kraev <egor.kraev@gmail.com>
Signed-off-by: whimo <whimox@gmail.com>
Signed-off-by: whimo <whimox@gmail.com>
Signed-off-by: Egor Kraev <egor.kraev@gmail.com>
Signed-off-by: Egor Kraev <egor.kraev@gmail.com>
Bump the core stack and adapt causaltune to the new APIs, preserving behavior (same estimator set, same scoring, same public API): - deps: dowhy 0.9.1->0.14, econml 0.14.1->0.16.0, numpy 1.23->2.2, FLAML 2.2->2.6, xgboost 1.7->2.1, sklearn <1.7, pandas <3; drop the setuptools pin; python_requires >=3.10,<3.13; CI matrix 3.10-3.12 - numpy 2: vendor is_sequence (numpy.distutils removed); np.trapz->np.trapezoid - dowhy 0.14: adapt DoWhyWrapper.fit/estimate_effect + CausalEstimate to the new signatures; effect_tt uniform 2-arg; rewrite effect_stderr on the native apply_multitreatment; treatment/outcome-name accessors for renamed adapter attributes; silence the string-dispatch deprecation - econml 0.16: hide predict_proba on FLAML regression outcome models so DR learners don't reject them; boolean search params for fit_intercept - sklearn 1.6 / shap 0.48 / dcor 0.7 / Python 3.12 fixups - add regression tests for each of the above Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: Egor Kraev <egor.kraev@gmail.com>
- lint (flake8): exclude notebooks/ (carried experiment runners) and add E226 to ignore (black-incompatible; pre-existing in datasets.py) - test_nhefs/lalonde/acic: skip on external-download failure (NHEFS host now serves HTML, not CSV) instead of failing CI - test_make_scores_all_backdoor_metrics: assert each metric is present and numeric rather than strictly finite (codec/frobenius/energy legitimately return inf on degenerate data; the point is exercising every metric branch) - params.py: IV defaults projection/opt_reweighted -> bool (sklearn 1.6 rejects int 0/1), matching the search-space fix (Codex) - scoring.py: forward N_values into bite() in bite_score (Codex) - CI: fail-fast: false so the matrix reports every OS/Python config Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: Egor Kraev <egor.kraev@gmail.com>
… fix) The test drew unseeded random data (global numpy RNG in linear_multi_dataset) and used no random_state on the split/LinearDML, with an over-tight 0.01 tolerance. It passed on 5/6 CI configs but drifted to 0.26 on windows-3.12. Seed the data generation, split, and estimator so the estimate is deterministic and reproducible across platforms; keep a 0.2 tolerance that reflects estimator variance rather than the flaky 0.01. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: Egor Kraev <egor.kraev@gmail.com>
The propensity model is fit on df[propensity_modifiers] (PropensityScoreWeighter) but scoring predicted on df[get_effect_modifiers() + get_common_causes()] -- a different list whose order is set-derived (_observed_common_causes_names) and thus nondeterministic across Python hash seeds. sklearn 1.6 now strictly enforces matching feature names AND order between fit and predict, so this flaked on windows/py3.10 (worsened by duplicate columns when effect_modifiers overlap common_causes). Predict on the exact list the model was fit on so the columns always match. Verified stable across 10 hash seeds; propensity/scoring suite (17 tests) unchanged. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: Egor Kraev <egor.kraev@gmail.com>
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
2 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.
What
Upgrade the core stack off the 2022-era pins and adapt CausalTune to the new APIs, preserving behavior (same estimator set, same scoring, same public API):
<1.7, pandas<3; drop thesetuptoolspin.python_requires>=3.10,<3.13); CI matrix and classifiers updated; runner + action versions bumped.Key adaptations
is_sequence(numpy.distutilsremoved);np.trapz→np.trapezoid.DoWhyWrapper.fit/estimate_effect+CausalEstimateconstruction to the new signatures; uniform 2-argeffect_tt; rewriteeffect_stderron the adapter's nativeapply_multitreatment;treatment_name_of/outcome_name_ofaccessors for the renamed adapter attributes; silence the string-dispatchDeprecationWarningnarrowly.predict_probaon FLAML regression outcome models so DR learners don't reject them as classifiers; boolean search params (fit_intercept/fit_cate_intercept).KernelExplainervs read-onlyfeature_names_in_;pairwise_distancestranspose inpsw_energy_distance;warnings.warncategory).Notes
main).Testing
Full non-slow suite green locally on the target stack, except
test_nhefs, which downloads an external CSV and only fails in the sandboxed environment (no outbound network to that host) — it should pass in CI.🤖 Generated with Claude Code