Migrate OutlierTrimmer to narwhals, add polars support - #1035
Open
solegalli wants to merge 2 commits into
Open
Conversation
Shared base for all outlier transformers (ArbitraryOutlierCapper extends BaseOutlier directly; Winsoriser/OutlierTrimmer extend WinsorizerBase): column reorder + NA/Inf checks in _check_transform_input_and_state(), the fold-limit estimation in WinsorizerBase.fit() (gaussian/iqr/mad/ quantiles), and the capping step in BaseOutlier._transform() are now dataframe-agnostic. Capping (np.clip against per-column bounds) was benchmarked three ways at 10k/50k/100k rows x 1/2/10 columns: pandas-native .clip() loop vs. a single narwhals with_columns(nw.col(v).clip(lo, hi) for v in ...) vs. grouping columns by which bound(s) apply and running up to 3 vectorized numpy calls (np.clip/minimum/maximum) via to_numpy()/new_series(), mirroring ReciprocalTransformer's numpy-acceleration pattern. narwhals-generic alone was already close to parity (0.95-1.49x pandas-native - minimal loss, mergeable per the imputation-base precedent), but the numpy-grouped version was faster still: 0.16-0.82x of pandas-native on the homogeneous case (single tail, all columns share the same bound - the common Winsoriser/ OutlierTrimmer case) and 0.42-1.52x on mixed-coverage dicts (the ArbitraryOutlierCapper case, up to 3 groups). Adopted the numpy-grouped version as the single merged code path for both backends. A first numpy attempt used a blanket -inf/inf sentinel for the missing side per column (like RelativeFeatures-style bound arrays) - that's a correctness bug, not just a style choice: mixing an int64 numpy array with a float -inf/inf bound upcasts the whole column to float64 even when the real, present bound is an int (e.g. ArbitraryOutlierCapper's own docstring example, `max_capping_dict=dict(x1=8)`, expects int64 out). Grouping columns into "both bounds" / "right only" / "left only" buckets and calling np.clip/minimum/maximum with only the bounds that actually exist avoids ever introducing an inf, so dtype promotion matches pandas .clip() exactly - verified byte-for-byte against the old pandas-only implementation across all 4 capping methods x 3 tails, plus the int-dtype and mixed-dict-coverage cases. Also found and fixed a real bug introduced while migrating fit(): plain np.mean/np.std/np.quantile/np.median propagate NaN, unlike pandas' mean/std/quantile/median which skip NaN by default. With missing_values="ignore" and NaN present, this silently produced NaN caps instead of the caps computed from non-null data. Fixed by using the nan-aware numpy variants (np.nanmean/nanstd/nanquantile/nanmedian). Caught by tests/test_outliers/test_winsorizer.py::test_transformer_ignores_na_in_df, which predates this migration but exercises exactly this path. variables/feature names can be int or str; passing a plain list to narwhals' .select() only works for string columns, so every .select() call here uses nw.col(*variables) instead - .select(list_of_ints) raises InvalidIntoExprError. Verified: tests/test_outliers full suite - 83 passed, 3 pre-existing failures in test_check_estimator_outliers.py (sklearn's check_estimator feeds raw numpy arrays, which check_X() has always rejected per the narwhals migration's dataframe-only contract; identical failure set before and after this change). flake8 and mypy clean on the file. Module imports and runs fit/_transform end-to-end on polars with pandas import fully blocked. sphinx -W build clean (only the pre-existing unrelated linkcode_resolve warning). All 4 capping-method x tail combinations and the Winsoriser/OutlierTrimmer/ArbitraryOutlierCapper docstring examples produce byte-identical output to the pre-migration code (checked exact numeric values and dtypes). Not migrated here (belongs to the 3 follow-on transformer branches): ArbitraryOutlierCapper.fit()/transform(), Winsoriser's add_indicators branch (pd.concat), and OutlierTrimmer.transform() (its own .le/.ge/.loc row-filtering, which doesn't go through BaseOutlier._transform at all) all still import pandas directly. Existing tests in tests/test_outliers were left pandas-only rather than parametrized over polars, since they exercise those still-pandas-only subclasses, not BaseOutlier/ WinsorizerBase directly - parametrizing them now would fail on reasons unrelated to this file. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
transform() now filters rows via a single narwhals .filter() call built from a combined boolean expression (AND of each variable's right/left cap conditions), instead of a pandas .loc masking loop. Benchmarked against pandas-native and a numpy boolean-mask extraction at 10k/50k/ 100k rows x 1/2/10 columns: the narwhals filter is within 1.4-1.75x of pandas-native at 10k rows (sub-millisecond absolute difference) and becomes faster than pandas-native from 50k rows up (0.58x-0.97x), so a single merged code path (no is_pandas branching) is the right call here - unlike BaseOutlier's elementwise capping, which benefits from numpy grouping, row-filtering is exactly what narwhals .filter() already pushes down to the native backend efficiently. Also fixes a latent bug in TransformXyMixin.transform_x_y() (_base_transformers/mixins.py): the non-pandas branch added a row-index marker column via with_row_index() and passed it straight to self.transform(), but never widened feature_names_in_/ n_features_in_ to account for it. Any transform() that validates column count (BaseOutlier._check_transform_input_and_state, via _check_X_matches_training_df) then raised a ValueError on the extra column. This was latent because no narwhals-migrated class on this branch previously combined TransformXyMixin with a column-count- checking transform() on a non-pandas backend - OutlierTrimmer is the first. The fix (guarded widen/restore of feature_names_in_ around the transform() call) is carried over verbatim from the same fix already applied to this file on branch narwhals-drop-missing-data (commit fd99caf), which hadn't been merged into this branch yet. Tests rewritten to one parametrized test per behavior over make_df in [pd.DataFrame, pl.DataFrame], plus a new test asserting that caps on two different variables combine with AND (each variable drops a distinct row) - a code path the old sequential-loop version exercised implicitly but no test isolated directly. Docs verified against live output: the class docstring's pandas examples were already accurate; the user guide's Titanic-based numbers had drifted from the current openml dataset (predates this migration, e.g. the IQR section's age max was already wrong against the old pandas-loop transform()) and are corrected here, plus a "With polars" section is added. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.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
1 participant
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.
Migrates
OutlierTrimmerto narwhals with polars support.transform()now filters rows via a single narwhals.filter()call built from a combined boolean expression (AND of each variable's right/left cap conditions), instead of a pandas.locmasking loop.Merge vs split: benchmarked against pandas-native and a numpy boolean-mask extraction at 10k/50k/100k rows × 1/2/10 cols. The narwhals filter is within 1.4–1.75x of pandas-native at 10k rows (sub-ms absolute) and faster from 50k rows up (0.58x–0.97x). Single merged path — row filtering is exactly what narwhals
.filter()pushes down to the native backend efficiently (unlikeBaseOutlier's elementwise capping, which benefits from numpy grouping).Bug fixed in
TransformXyMixin.transform_x_y()(_base_transformers/mixins.py): the non-pandas branch added a row-index marker column viawith_row_index()and passed it toself.transform(), but never widenedfeature_names_in_/n_features_in_, so a column-count-checkingtransform()raisedValueErroron the extra column. Latent becauseOutlierTrimmeris the first narwhals-migrated class to combineTransformXyMixinwith a column-count-checkingtransform()on a non-pandas backend. Fix (guarded widen/restore around thetransform()call) carried over verbatim from the same fix onnarwhals-drop-missing-data(commit fd99caf), not yet merged into this branch.Tests rewritten to one parametrized test per behaviour over
make_df in [pd.DataFrame, pl.DataFrame], plus a new test asserting caps on two different variables combine with AND. Docs: user-guide Titanic numbers had drifted from the current openml dataset (predates this migration) — corrected here, "With polars" section added.Stacked on
narwhals-outliers-base(its own PR). Until that merges this PR's diff also contains the sharedBaseOutlier/WinsorizerBasecommit; review that one first.