[dask, sklearn] Fix predict proba. (#6566)

* For sklearn:
  - Handles user defined objective function.
  - Handles `softmax`.

* For dask:
  - Use the implementation from sklearn, the previous implementation doesn't perform any extra handling.

python-package/xgboost/dask.py

@@ -40,6 +40,7 @@ from .training import train as worker_train
 from .tracker import RabitTracker, get_host_ip
 from .sklearn import XGBModel, XGBRegressorBase, XGBClassifierBase, _objective_decorator
 from .sklearn import xgboost_model_doc
+from .sklearn import _cls_predict_proba
 
 
 if TYPE_CHECKING:
@@ -1504,6 +1505,10 @@ class DaskXGBClassifier(DaskScikitLearnBase, XGBClassifierBase):
                               early_stopping_rounds=early_stopping_rounds,
                               callbacks=callbacks)
         self._Booster = results['booster']
+
+        if not callable(self.objective):
+            self.objective = params["objective"]
+
         # pylint: disable=attribute-defined-outside-init
         self.evals_result_ = results['history']
         return self
@@ -1554,7 +1559,7 @@ class DaskXGBClassifier(DaskScikitLearnBase, XGBClassifierBase):
                                    data=test_dmatrix,
                                    validate_features=validate_features,
                                    output_margin=output_margin)
-        return pred_probs
+        return _cls_predict_proba(self.objective, pred_probs, da.vstack)
 
     # pylint: disable=arguments-differ,missing-docstring
     def predict_proba(
@@ -1593,6 +1598,8 @@ class DaskXGBClassifier(DaskScikitLearnBase, XGBClassifierBase):
             output_margin=output_margin,
             validate_features=validate_features
         )
+        if output_margin:
+            return pred_probs
 
         if self.n_classes_ == 2:
             preds = (pred_probs > 0.5).astype(int)

python-package/xgboost/sklearn.py

@@ -819,6 +819,20 @@ class XGBModel(XGBModelBase):
         return np.array(json.loads(b.get_dump(dump_format='json')[0])['bias'])
 
 
+def _cls_predict_proba(objective: Union[str, Callable], prediction: Any, vstack: Callable) -> Any:
+    if objective == 'multi:softmax':
+        raise ValueError('multi:softmax objective does not support predict_proba,'
+                         ' use `multi:softprob` or `binary:logistic` instead.')
+    if objective == 'multi:softprob' or callable(objective):
+        # Return prediction directly if if objective is defined by user since we don't
+        # know how to perform the transformation
+        return prediction
+    # Lastly the binary logistic function
+    classone_probs = prediction
+    classzero_probs = 1.0 - classone_probs
+    return vstack((classzero_probs, classone_probs)).transpose()
+
+
 @xgboost_model_doc(
     "Implementation of the scikit-learn API for XGBoost classification.",
     ['model', 'objective'], extra_parameters='''
@@ -929,7 +943,9 @@ class XGBClassifier(XGBModel, XGBClassifierBase):
                               verbose_eval=verbose, xgb_model=model,
                               callbacks=callbacks)
 
-        self.objective = params["objective"]
+        if not callable(self.objective):
+            self.objective = params["objective"]
+
         if evals_result:
             for val in evals_result.items():
                 evals_result_key = list(val[1].keys())[0]
@@ -1031,7 +1047,8 @@ class XGBClassifier(XGBModel, XGBClassifierBase):
         Returns
         -------
         prediction : numpy array
-            a numpy array with the probability of each data example being of a given class.
+            a numpy array of shape array-like of shape (n_samples, n_classes) with the
+            probability of each data example being of a given class.
         """
         test_dmatrix = DMatrix(X, base_margin=base_margin,
                                missing=self.missing, nthread=self.n_jobs)
@@ -1040,11 +1057,7 @@ class XGBClassifier(XGBModel, XGBClassifierBase):
         class_probs = self.get_booster().predict(test_dmatrix,
                                                  ntree_limit=ntree_limit,
                                                  validate_features=validate_features)
-        if self.objective == "multi:softprob":
+        return _cls_predict_proba(self.objective, class_probs, np.vstack)
-            return class_probs
-        classone_probs = class_probs
-        classzero_probs = 1.0 - classone_probs
-        return np.vstack((classzero_probs, classone_probs)).transpose()
 
     def evals_result(self):
         """Return the evaluation results.

tests/python/test_with_dask.py

@@ -160,7 +160,7 @@ def test_boost_from_prediction(tree_method: str) -> None:
                 tree_method=tree_method,
             )
             model_0.fit(X=X_, y=y_)
-            margin = model_0.predict_proba(X_, output_margin=True)
+            margin = model_0.predict(X_, output_margin=True)
 
             model_1 = xgb.dask.DaskXGBClassifier(
                 learning_rate=0.3,

tests/python/test_with_sklearn.py

@@ -79,6 +79,18 @@ def test_multiclass_classification():
         check_pred(preds3, labels, output_margin=True)
         check_pred(preds4, labels, output_margin=False)
 
+    cls = xgb.XGBClassifier(n_estimators=4).fit(X, y)
+    assert cls.n_classes_ == 3
+    proba = cls.predict_proba(X)
+    assert proba.shape[0] == X.shape[0]
+    assert proba.shape[1] == cls.n_classes_
+
+    # custom objective, the default is multi:softprob so no transformation is required.
+    cls = xgb.XGBClassifier(n_estimators=4, objective=tm.softprob_obj(3)).fit(X, y)
+    proba = cls.predict_proba(X)
+    assert proba.shape[0] == X.shape[0]
+    assert proba.shape[1] == cls.n_classes_
+
 
 def test_ranking():
     # generate random data
@@ -788,6 +800,11 @@ def test_save_load_model():
         booster.save_model(model_path)
         cls = xgb.XGBClassifier()
         cls.load_model(model_path)
+
+        proba = cls.predict_proba(X)
+        assert proba.shape[0] == X.shape[0]
+        assert proba.shape[1] == 2  # binary
+
         predt_1 = cls.predict_proba(X)[:, 1]
         assert np.allclose(predt_0, predt_1)
 

tests/python/testing.py

@@ -253,6 +253,34 @@ def eval_error_metric(predt, dtrain: xgb.DMatrix):
     return 'CustomErr', np.sum(r)
 
 
+def softmax(x):
+    e = np.exp(x)
+    return e / np.sum(e)
+
+
+def softprob_obj(classes):
+    def objective(labels, predt):
+        rows = labels.shape[0]
+        grad = np.zeros((rows, classes), dtype=float)
+        hess = np.zeros((rows, classes), dtype=float)
+        eps = 1e-6
+        for r in range(predt.shape[0]):
+            target = labels[r]
+            p = softmax(predt[r, :])
+            for c in range(predt.shape[1]):
+                assert target >= 0 or target <= classes
+                g = p[c] - 1.0 if c == target else p[c]
+                h = max((2.0 * p[c] * (1.0 - p[c])).item(), eps)
+                grad[r, c] = g
+                hess[r, c] = h
+
+        grad = grad.reshape((rows * classes, 1))
+        hess = hess.reshape((rows * classes, 1))
+        return grad, hess
+
+    return objective
+
+
 class DirectoryExcursion:
     def __init__(self, path: os.PathLike, cleanup=False):
         '''Change directory.  Change back and optionally cleaning up the directory when exit.