mapie/removed-in-mapie
| Rule | removed-in-mapie |
|---|---|
| Pack | MAPIE |
| Severity | error |
| Fix | unsafe, with xpd check --fix --unsafe-fixes |
| Library versions | MAPIE 1.0 and later |
The MAPIE 0.x API was removed, so the import or the call raises on MAPIE 1.6. Use the 1.x class that the message names.
from mapie.regression import MapieRegressorfrom mapie.metrics import regression_coverage_scorefrom mapie.regression import SplitConformalRegressorfrom mapie.metrics.regression import regression_coverage_score
mapie = MapieRegressor(model, cv="prefit")mapie.fit(X_conf, y_conf)y_pred, y_pis = mapie.predict(X_test, alpha=0.1)coverage = regression_coverage_score(y_test, y_pis[:, 0, 0], y_pis[:, 1, 0])mapie = SplitConformalRegressor(model, confidence_level=0.9, prefit=True)mapie.conformalize(X_conf, y_conf)y_pred, y_pis = mapie.predict_interval(X_test)coverage = regression_coverage_score(y_test, y_pis)[0]What the rule reports
Section titled “What the rule reports”The rule reports the MAPIE 0.x API that raises on MAPIE 1.6.0:
| Old form | Where the rule reports it |
|---|---|
MapieRegressor, MapieClassifier, MapieQuantileRegressor, MapieTimeSeriesRegressor, MapieCalibrator, MapieMultiLabelClassifier, PrecisionRecallController, MondrianCP, ConformityScore |
an import |
modules that are gone, such as mapie.quantile_regression, mapie.control_risk and mapie._typing |
an import |
metric functions from mapie.metrics itself |
an import |
alpha= |
the MAPIE 1 constructors, predict_interval, predict_set, predict of an object that the file builds from a MAPIE 1 class, and the metrics |
fit |
a CrossConformalRegressor, CrossConformalClassifier or JackknifeAfterBootstrapRegressor |
cv= |
the split classes |
cv="prefit" or cv="split" |
the cross classes |
method= |
the classifiers and TopLabelCalibrator |
method="naive" |
any class |
sample_weight= |
fit or fit_conformalize |
partial_fit |
a TimeSeriesRegressor |
| the lower and upper bounds as two arguments | a metric |
The rule leaves alone an import inside a try block that catches the error, code behind a version
check, and code inside pytest.raises. Each of its documents carries the release that removed the
form:
| Form | Removed in |
|---|---|
| most forms | 1.0 |
partial_fit |
1.2 |
PrecisionRecallController |
1.3 |
sample_weight= on TimeSeriesRegressor.fit |
1.4.1 |
How to fix it
Section titled “How to fix it”xpd check --fix --unsafe-fixes applies the fixes of this rule. Check the result, because an unsafe fix can change what the code does.
The safe fix writes numpy.typing for mapie._typing, because on MAPIE 0.9.2 ArrayLike and NDArray of mapie._typing are the objects of numpy.typing. The unsafe fix writes mapie.metrics.regression, mapie.metrics.classification or mapie.metrics.calibration for mapie.metrics when every imported name is in that module, confidence_level=1 - alpha for alpha= on a constructor, on TimeSeriesRegressor.predict, on regression_mwi_score and on coverage_width_based when the value is a number or a name, fit_params={"sample_weight": w} for sample_weight=w, update for partial_fit, and one stacked array for the two bounds given to regression_coverage_score and regression_mean_width_score. A list of numbers given as alpha= on a constructor or on TimeSeriesRegressor.predict becomes confidence_level= with 1 minus each number, fit_conformalize replaces fit on a cross or jackknife object, method= becomes conformity_score= on a classifier when its value is "lac", "aps", "raps" or "top_k", and method= is taken out of TopLabelCalibrator, which drops the comments between the arguments. The removed classes and modules, alpha= on predict_interval, predict_set and predict, cv= and the other values of method have no fix, because the rewrite reaches other calls or the message names more than one. No release ran an unsafe fix's old and new form side by side, except sample_weight= and fit_params= on TimeSeriesRegressor.fit in 1.4.0, which ignored sample_weight= and gave other values, and partial_fit and update in 1.1.0, which are two methods.
The forms around partial_fit and sample_weight changed in the following releases:
| Form | Worked on | Failed on |
|---|---|---|
TimeSeriesRegressor.partial_fit |
1.1.0, in the wheel | 1.2.0, where it is gone |
sample_weight= on TimeSeriesRegressor.fit |
up to 1.4.0 | 1.4.1 and later, with TypeError |
fit_params={"sample_weight": w} |
1.4.0 and later | 1.0.1 to 1.3.0 |
sample_weight= on fit of the split classes |
not given | 1.0.1 already |
The fix confidence_level=1 - alpha is written only for a number or a name. alpha=[0.05, 0.2]
was common in 0.x code, and 1 - [0.05, 0.2] raises TypeError, so a list of numbers gets
confidence_level=[1 - 0.05, 1 - 0.2]. Another expression is reported without a fix, because the
file does not show whether it is one number or a list. confidence_level=[0.95, 0.8] gave
intervals of shape (3, 2, 2) for 3 rows.
predict(X, alpha=0.1) raised TypeError on each of the split, cross, jackknife and quantile
classes. Their level is set in the constructor, so the predict document has no fix. fit raised
AttributeError on CrossConformalRegressor, CrossConformalClassifier and
JackknifeAfterBootstrapRegressor, on 1.0.1 and on 1.6.0. The rule reports it on a name that the
file assigns from one of the three and from nothing else. CrossConformalizedQuantileRegressor has
a fit on 1.6.0 and is left alone.
The rule reports an import and doesn't rename the class, because the constructors differ.
MapieRegressor(model, cv="prefit") becomes SplitConformalRegressor(model, prefit=True), and
alpha moves from predict to the constructor. Two fixes write np. and apply only in a file
with import numpy as np. On the same bounds, the fixed calls of regression_coverage_score,
regression_mean_width_score and coverage_width_based on MAPIE 1.6.0 returned the value that
the old calls returned on MAPIE 0.9.2, bit for bit, and regression_mwi_score returned it to 15
digits.
Message
Section titled “Message”xpd prints this message with the issue. A name that starts with $ stands for code from your file.
MapieRegressor was removed in MAPIE 1.0, and the import raises ImportError. Use SplitConformalRegressor for cv="split" and cv="prefit", CrossConformalRegressor for a number of folds or a splitter, and JackknifeAfterBootstrapRegressor for a Subsample. Pass confidence_level, which is 1 - alpha, to the constructor, and read the intervals from predict_interval.Browse the other rules of the MAPIE pack.