Add "full" strategy to NCH (#353)

* Update classification.py

* [pre-commit.ci] auto fixes from pre-commit.com hooks

* Update classification.py

* [pre-commit.ci] auto fixes from pre-commit.com hooks

* Update test_classification.py

* [pre-commit.ci] auto fixes from pre-commit.com hooks

* Update pyriemann_qiskit/classification.py

Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com>

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com>

Author: 3 people

pyriemann_qiskit/classification.py

@@ -867,11 +867,14 @@ class NearestConvexHull(BaseEstimator, ClassifierMixin, TransformerMixin):
         The number of hulls used per class, when subsampling is "random".
     n_samples_per_hull : int, default=15
         Defines how many samples are used to build a hull. -1 will include
-        all samples per class.
-    subsampling : {"min", "random"}, default="min"
-        Subsampling strategy of training set to estimate distance to hulls.
-        "min" estimates hull using the n_samples_per_hull closest matrices.
-        "random" estimates hull using n_samples_per_hull random matrices.
+        all samples per class. If subsampling is "full", this
+        parameter is defaulted to -1.
+    subsampling : {"min", "random", "full"}, default="min"
+        Subsampling strategy of training set to estimate distance to hulls:
+
+        - "min" estimates hull using the n_samples_per_hull closest matrices;
+        - "random" estimates hull using n_samples_per_hull random matrices;
+        - "full" estimates the hull using the entire training matrices, as in [1]_.
     seed : float, default=None
         Optional random seed to use when subsampling is set to `random`.
 
@@ -900,9 +903,14 @@ def __init__(
         self.subsampling = subsampling
         self.seed = seed
 
-        if subsampling not in ["min", "random"]:
+        if subsampling not in ["min", "random", "full"]:
             raise ValueError(f"Unknown subsampling type {subsampling}.")
 
+        if subsampling == "full":
+            # From code perspective, "full" strategy is the same as min strategy
+            # without sorting
+            self.n_samples_per_hull = -1
+
     def fit(self, X, y):
         """Fit (store the training data).
 
@@ -996,7 +1004,7 @@ def _predict_distances(self, X):
         if self.debug:
             print("Total test samples:", X.shape[0])
 
-        if self.subsampling == "min":
+        if self.subsampling == "min" or self.subsampling == "full":
             self._process_sample = self._process_sample_min_hull
         elif self.subsampling == "random":
             self._process_sample = self._process_sample_random_hull

tests/test_classification.py

@@ -61,6 +61,27 @@ def test_qsvm_init(quantum):
     assert not hasattr(q, "_provider")
 
 
+def test_nch_init_full():
+    """Test init of NCH classifier with `full` subsampling"""
+
+    q = QuanticNCH(subsampling="full").fit(X=np.array([[0], [1]]), y=np.array([0, 1]))
+    clf = q._classifier
+
+    assert type(clf).__name__ == "NearestConvexHull"
+
+    # Check "full" subsampling takes all training samples.
+    assert clf.n_samples_per_hull == -1
+
+    # Check it calls the "min" strategy on prediction.
+    def mocked_process_sample_min_hull(*args):
+        raise ValueError("Min strategy called")
+
+    clf._process_sample_min_hull = mocked_process_sample_min_hull
+
+    with pytest.raises(ValueError):
+        clf._predict_distances(X=[0])
+
+
 class TestQSVMSplitClasses(BinaryTest):
     """Test _split_classes method of quantum classifiers"""