yzhao062 / pyod / 25717592359

94%

master: 94%

Build Type

push

github

Committed by web-flow

Commit Message

fix(lunar): stop sharing scaler state across instances (closes #502) (#674)

`LUNAR.__init__` had `scaler=MinMaxScaler()` as a default argument,
which Python evaluates once at module import time. Two `LUNAR()`
instances therefore shared the same scaler object; the second `.fit`
re-fit it under a different feature dimensionality and silently
invalidated the first instance's `predict` (issue #502 reproducer
crashes with a shape error on the third `predict` call).

The fix:
- Default `scaler` to `None` and store the constructor arg verbatim
  on `self.scaler` (sklearn.base.clone requires that __init__ does not
  mutate its inputs — verified by the existing `test_model_clone`).
- Resolve a per-fit scaler in a new `_resolve_scaler` helper called
  from `fit`. The fitted scaler lives on `self.scaler_` (sklearn
  fitted-attribute convention); a fresh `MinMaxScaler()` is created
  when `scaler is None`, scaling is disabled when `scaler is False`,
  and any other instance is deep-copied so two LUNAR instances
  configured with the same scaler object cannot trample each other.
- Update `decision_function` and the train-time score path to read
  `self.scaler_` instead of `self.scaler`.

Two regression tests (`TestLUNARScalerIsolation`) reproduce the
original failure mode on master and pass on this branch.

AI disclosure: drafted with assistance from Claude (Anthropic).

Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>

Coverage Stats

42 of 43 new or added lines in 2 files covered. (97.67%)

21 existing lines in 7 files now uncovered.

17905 of 19098 relevant lines covered (93.75%)

10.29 hits per line