Migrate EqualFrequencyDiscretiser.fit() to narwhals, add polars support - #1039
Open
solegalli wants to merge 2 commits into
Open
Migrate EqualFrequencyDiscretiser.fit() to narwhals, add polars support#1039solegalli wants to merge 2 commits into
solegalli wants to merge 2 commits into
Conversation
Shared base for ArbitraryDiscretiser, EqualFrequencyDiscretiser,
EqualWidthDiscretiser and GeometricWidthDiscretiser (not
DecisionTreeDiscretiser, which extends a different base). Only
transform() needed migrating - _fit_setup(), _get_feature_names_in()
and _check_transform_input_and_state() are inherited unchanged from
BaseNumericalTransformer, already fully narwhals-migrated.
transform()'s only pandas dependency was pd.cut, applied per column to
sort values into the bins already fixed by fit() (binner_dict_).
Replaced it with a plain numpy implementation: pandas.cut is itself
built on bins.searchsorted() internally (verified against pandas 3.0's
_bins_to_cuts source), so np.searchsorted + the same include_lowest
index-1 special case reproduces its bin-index logic exactly, with no
per-backend branch needed - values come from
nw_X.get_column(feature).to_numpy() regardless of backend, and results
are re-attached via nw.new_series()/with_columns(), so the same code
path runs for pandas and polars.
Benchmarked old pd.cut vs the new numpy+narwhals path at 10k/50k/100k
rows x 1/2/10 columns:
- return_boundaries=False (bin codes): narwhals-on-pandas lands at
~1.0-1.2x of pandas-native at realistic sizes (50k-100k rows, the
~1.9x seen only at the smallest 10k-row/1-col case is fixed
per-call overhead, sub-millisecond either way) - minimal loss,
merged into a single path, no is_pandas split. narwhals-on-polars is
~1.0-1.3x *faster* than pandas-native at every size tested.
- return_boundaries=True (interval-label strings): the numpy path is
12-20x faster than pd.cut on pandas itself (e.g. 100k rows x 10
cols: 647ms old vs 40ms new) - pd.cut's Categorical/IntervalIndex
machinery has heavy per-call overhead that np.searchsorted plus
plain string formatting avoids entirely. polars is ~1.2x faster
still than the new pandas path.
Given both branches favour or are at parity with a single numpy-driven
path, there was no case for a pandas fast-path split here.
return_boundaries=True's interval-label formatting
("(lower, upper]" text, e.g. "(-0.001, 20.0]") replicates pandas.cut's
_round_frac/_infer_precision/lowest-edge-adjustment algorithm in pure
numpy so it works identically on both backends - verified against real
pd.cut(...).astype(str) output across positive/negative/duplicate-
inducing/inf-edge bins, and against the California housing dataset
used in the existing test. return_object=True now builds a nw.Object
column (narwhals' cross-backend equivalent of pandas' "O" dtype,
already used by variable_handling for categorical-column detection)
instead of a pandas-only astype("O") call.
Verified: tests/test_discretisation full suite unchanged (109 passed,
5 pre-existing failures in test_check_estimator_discretisers.py -
sklearn's check_estimator feeds raw numpy arrays, which check_X() has
rejected since the narwhals migration's dataframe-only contract;
reproduced identically on the unmodified file). Manually diffed
transform() output against real pd.cut() across ~10 edge cases (NaN,
out-of-range values on both ends, negative bins, exact-edge values,
precision auto-widening, single bin) plus the three sibling
discretisers' documented doctest examples (EqualWidthDiscretiser,
ArbitraryDiscretiser, EqualFrequencyDiscretiser value_counts()) -
all numerically identical to old pd.cut output; the "Name: x" vs
"Name: count" and bare-fit()-repr mismatches those doctests already
show are a pre-existing pandas-3.0 doc-staleness issue unrelated to
this migration (reproduced on the unmodified files too). flake8 and
mypy clean. Module imports with pandas blocked (loaded standalone,
since sibling discretiser files in this package are not yet migrated
and still import pandas at their own module level). sphinx -W build
clean (only the pre-existing unrelated linkcode_resolve warning).
test_base_discretizer.py's test_transform is now parametrized over
pd.DataFrame/pl.DataFrame per AGENTS.md - its MockClassFit hard-codes
binner_dict_ rather than actually fitting, so it needed no pandas-only
logic to begin with. The other four discretisers' own test files stay
pandas-only for now: their fit() methods still call pd.cut/pd.qcut
directly and aren't migrated by this branch.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
fit()'s only pandas dependency was pd.qcut(duplicates="drop"), used to
compute quantile-based bin edges per variable. Replaced it with
np.quantile() on each column's narwhals-extracted numpy array, plus
np.unique() to sort and drop duplicate edges - reproducing qcut's
duplicates="drop" behaviour without any per-backend branch, since
values come from nw_X.get_column(var).to_numpy() regardless of
backend.
Getting a bit-exact match (not just numerically close) took two fixes
verified against pandas 3.0's pandas.core.reshape.tile.qcut source:
- pandas masks out NaN before calling np.quantile(values, qs,
method="linear") itself, rather than using np.nanquantile - the two
are not always bit-identical. Here this distinction is moot in
practice: _fit_setup() already rejects NaN in variables_, so no
masking is needed - values reaching the loop are already NaN-free.
- qcut nudges each quantile that isn't exactly representable in base 2
up via np.nextafter (np.linspace(0, 1, q+1) then
np.putmask(quantiles, q*quantiles != np.arange(q+1),
nextafter(quantiles, 1))), rounding up rather than to nearest.
Skipping this shifted bin edges by ~1e-13 versus real pd.qcut
output and broke an existing exact-equality test.
With both applied, verified bit-exact (np.array_equal) against real
pd.qcut(retbins=True) across large random floats, many-duplicate-value
data, all-identical-value data, negative floats, and n<q data.
Benchmarked old pd.qcut vs the new numpy+narwhals path at 10k/50k/100k
rows x 1/2/10 columns: the new path is consistently faster than the
old pandas-native code on BOTH backends (narwhals-on-pandas lands at
0.19x-0.47x of old pd.qcut's time, narwhals-on-polars at 0.12x-0.46x,
both converging to roughly 2x faster at realistic 50k-100k row sizes).
A narwhals-native quantile-expression alternative was also benchmarked
(one nw.col(var).quantile(qi) expr per quantile point, batched into a
single select()) - fast on polars but 2-3x *slower* than old pd.qcut
on pandas, since narwhals translates each expr to a separate
Series.quantile call there. Given the numpy path beats old pandas on
both backends, there was no case for a pandas fast-path split.
Verified: tests/test_discretisation full suite unchanged (114 passed,
5 pre-existing failures in test_check_estimator_discretisers.py,
reproduced identically on the unmodified branch tip - sklearn's
check_estimator feeds raw numpy arrays, rejected since the narwhals
migration's dataframe-only contract). flake8 and mypy clean. Module
imports with pandas blocked (loaded standalone, since sibling
discretiser files in this package aren't migrated yet). sphinx -W
build clean (only the pre-existing unrelated linkcode_resolve
warning).
test_equal_frequency_discretiser.py rewritten per AGENTS.md: each
behaviour is now one test parametrized over
@pytest.mark.parametrize("make_df", [pd.DataFrame, pl.DataFrame])
rather than pandas-only.
docs/user_guide/discretisation/EqualFrequencyDiscretiser.rst: verified
every code example against real current output. The `disc.binner_dict_`
printout had two stale float digits (8099.200000000003 ->
...004, 1601.6000000000001 -> ...004, 1717.6999999999998 ->
1717.7000000000003) - reproduced identically with the OLD pd.qcut-based
fit() on the same dataset/pandas version, so this predates the
migration and is a doc-staleness issue, not a regression. Also
corrected the "uses pandas.qcut() under the hood" line and added a
"With polars" section with a verified worked example.
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
EqualFrequencyDiscretiser.fit()to narwhals with polars support.fit()'s only pandas dependency waspd.qcut(duplicates="drop")for quantile-based bin edges per variable. Replaced withnp.quantile()on each column's narwhals-extracted numpy array +np.unique()to sort and drop duplicate edges — reproducingqcut'sduplicates="drop"without any per-backend branch.Getting a bit-exact match (not just close) took two fixes verified against pandas 3.0's
qcutsource:np.quantile(values, qs, method="linear")itself rather than usingnp.nanquantile(not always bit-identical). Moot here —_fit_setup()already rejects NaN invariables_.qcutnudges each quantile not exactly representable in base 2 up vianp.nextafter(rounding up, not to nearest). Skipping this shifted edges by ~1e-13 and broke an existing exact-equality test.With both applied, verified
np.array_equalagainst realpd.qcut(retbins=True)across large random floats, many-duplicate data, all-identical data, negative floats, and n<q data.Merge vs split: benchmarked old
pd.qcutvs the new numpy+narwhals path at 10k/50k/100k rows × 1/2/10 cols — the new path is consistently faster than the old pandas-native code on both backends (narwhals-on-pandas 0.19x–0.47x of oldpd.qcut, narwhals-on-polars 0.12x–0.46x). A narwhals-native quantile-expression alternative was 2–3x slower than oldpd.qcuton pandas. No case for a split.Verified:
tests/test_discretisationunchanged (114 passed, 5 pre-existingcheck_estimatorfailures, reproduced on the unmodified branch tip). flake8 / mypy clean, sphinx -W clean.test_equal_frequency_discretiser.pyrewritten to one parametrized test per behaviour over[pd.DataFrame, pl.DataFrame].EqualFrequencyDiscretiser.rstexamples verified against real output (two stale float digits in thebinner_dict_printout reproduce with the oldpd.qcutfit — doc staleness, not a regression); "uses pandas.qcut() under the hood" line corrected, "With polars" section added.Stacked on
narwhals-discretisation-base(its own PR). Until that merges this PR's diff also contains the sharedBaseDiscretisercommit; review that one first.