Merge branch 'master' into check-regul-scaling

mislabeled/detect/detectors.py

@@ -21,7 +21,7 @@
     ProgressiveEnsemble,
 )
 from mislabeled.probe import (
-    ALOO,
+    ApproximateLOO,
     Confidence,
     CrossEntropy,
     FiniteDiffSensitivity,
@@ -70,7 +70,7 @@ def __init__(self, base_model):
         )
 
 
-class ALOODetector(ModelProbingDetector):
+class ApproximateLOODetector(ModelProbingDetector):
     """Detector based on Approximate LeaveOneOut for GLM models.
 
     References
@@ -85,7 +85,7 @@ def __init__(self, base_model):
         super().__init__(
             base_model=base_model,
             ensemble=NoEnsemble(),
-            probe=ALOO(),
+            probe=ApproximateLOO(),
             aggregate="sum",
         )
 

mislabeled/probe/__init__.py

@@ -16,7 +16,7 @@
 from ._complexity import ParameterCount, ParamNorm2
 from ._grads import GradSimilarity
 from ._influence import (
-    ALOO,
+    ApproximateLOO,
     GradNorm2,
     Representer,
     SelfInfluence,
@@ -47,7 +47,7 @@
     "ParameterCount",
     "ParamNorm2",
     "SelfInfluence",
-    "ALOO",
+    "ApproximateLOO",
     "Representer",
     "GradNorm2",
     "GradSimilarity",

mislabeled/probe/_influence.py

@@ -37,7 +37,7 @@ def __call__(self, estimator, X, y):
         return self_influence
 
 
-class ALOO(Maximize):
+class ApproximateLOO(Maximize):
     def __init__(self):
         pass
 

mislabeled/probe/_linear.py

@@ -152,6 +152,7 @@ def pseudo(self, X):
         return X
 
     def variance(self, p):
+        # variance of the GLM link function
         if self.loss == "l2":
             return np.eye(self.out_dim)[None, :, :] * np.ones(p.shape[0])[:, None, None]
         elif self.loss == "log_loss":

tests/probe/test_linear.py

@@ -88,7 +88,7 @@ def f(prc):
         print(np.round(jacobian(vectorized_objective, packed_raveled_coef).df, 2))
 
     # I dont know why the gradient should not take into account the regul
-    # to compute ALOO and SelfInfluence ...
+    # to compute ApproximateLOO and SelfInfluence ...
     # np.testing.assert_allclose(
     #     linearized.grad_p(X, y).sum(axis=0),
     #     jacobian(vectorized_objective, packed_raveled_coef).df,

tests/probe/test_self_influence.py

@@ -28,7 +28,7 @@
 from statsmodels.genmod import families
 from statsmodels.genmod.generalized_linear_model import GLM
 
-from mislabeled.probe import ALOO, SelfInfluence, linearize
+from mislabeled.probe import ApproximateLOO, SelfInfluence
 
 
 @pytest.mark.parametrize(
@@ -92,7 +92,7 @@ def loss_fn(model, X, y):
     model.fit(X, y)
 
     si = SelfInfluence()
-    aloo = ALOO()
+    aloo = ApproximateLOO()
 
     si_scores = si(model, X, y)
     aloo_scores = aloo(model, X, y)

tests/test_detect.py

@@ -22,6 +22,7 @@
 from mislabeled.aggregate import oob, sum
 from mislabeled.detect import ModelProbingDetector
 from mislabeled.detect.detectors import (
+    ApproximateLOODetector,
     AreaUnderMargin,
     Classifier,
     ConfidentLearning,
@@ -165,6 +166,12 @@ def simple_detect_test(n_classes, detector):
         )
     ),
     VoLG(MLPClassifier(random_state=seed)),
+    ApproximateLOODetector(
+        make_pipeline(
+            Nystroem(gamma=0.1, n_components=100, random_state=seed),
+            LogisticRegression(random_state=seed, C=10),
+        )
+    ),
 ]