Multiclass classification (#178)

* Update conftest.py

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

* Update test_classification.py

add multilabel test for quantum vqc

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

* Update conftest.py

change get_binary_feat to get_separable_feat

* Update test_classification.py

* Update test_classification.py

* Update test_classification.py

missing return

* Update conftest.py

replace base by super

* Update conftest.py

typo

* Update classification.py

renamed:
- _map_0_1_to_classes
- _map_1_0_to_classes

remove exception for binary classification in :
- _split_classes and fit

* Update test_classification.py

* Update Dockerfile

* Update Dockerfile

* Update classification.py

change class to idx and split methods to support multi labels classification

* Update test_classification.py

add test for split classes with n_classes = 3

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

* Update test_classification.py

fix test on split_classes

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

* Update test_classification.py

typo

* Update classification.py

fix bug with split_classes (index error)

* Update test_classification.py

typo

* Update test_classification.py

typo

* Update conftest.py

remove hardcoded n_classes=2

* Update test_classification.py

fix the number of sample

* Update conftest.py

fix class_len not updated after updating number of classes from 2 to 3

* Update classification.py

general one-hot encoded for multi-class

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

* Update classification.py

typo

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

* Update test_classification.py

flake8

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

* Update utils.py

changed get mne dataset to return auditory stimulation if required

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

* Create plot_multiclass_classification.py

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

* Update plot_multiclass_classification.py

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

* Update plot_multiclass_classification.py

* Update plot_multiclass_classification.py

fix imports and display

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

* - remove custom implementation of one-hot encoding (already done within Qiskit now)
- remove multilabels->multiclasses
- complete example

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

* rename Multiclasses -> Multiclass

* inverted y_pred and y_true in balanced accuracy

* Testing the code. Comments have been updated.

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

* Updated comments.

* Update conftest.py

* Update plot_classify_EEG_quantum_svm.py

* Update utils.py

* Update comments in classification.py

* Update multiclass_classification.py

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

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Gregoire Cattan <gregoire.cattan@ibm.com>
Co-authored-by: toncho11 <toncho11@users.noreply.github.com>

examples/ERP/multiclass_classification.py

@@ -0,0 +1,124 @@
+"""
+====================================================================
+Multiclass EEG classification with Quantum Pipeline
+====================================================================
+
+This example demonstrates multiclass EEG classification with a quantum
+classifier.
+We will be comparing the performance of VQC vs Quantum SVM vs
+Classical SVM vs Quantum MDM vs MDM.
+Execution takes approximately 1h.
+"""
+# Author: Gregoire Cattan
+# Modified from plot_classify_EEG_quantum_svm
+# License: BSD (3-clause)
+
+from pyriemann_qiskit.datasets import get_mne_sample
+from pyriemann_qiskit.pipelines import (
+    QuantumClassifierWithDefaultRiemannianPipeline,
+    QuantumMDMWithRiemannianPipeline,
+)
+from pyriemann.estimation import ERPCovariances
+from pyriemann.classification import MDM
+from sklearn.pipeline import make_pipeline
+from sklearn.model_selection import train_test_split
+from sklearn.metrics import (
+    confusion_matrix,
+    ConfusionMatrixDisplay,
+    balanced_accuracy_score,
+)
+from sklearn.decomposition import PCA
+from matplotlib import pyplot as plt
+
+
+print(__doc__)
+
+###############################################################################
+# Get the data
+
+# Use MNE sample. The include_auditory parameter select 4 classes.
+X, y = get_mne_sample(n_trials=-1, include_auditory=True)
+
+# evaluation without k-fold cross-validation
+X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3, random_state=1)
+
+###############################################################################
+# Decoding in tangent space with a quantum classifier
+
+# Our helper class QuantumClassifierWithDefaultRiemannianPipeline allows to
+# auto-configure the parameters of the pipelines.
+# Warning: these are not optimal parameters
+
+# Piepeline 1
+quantum_svm = QuantumClassifierWithDefaultRiemannianPipeline(
+    dim_red=PCA(n_components=5),
+)
+
+# Piepeline 2
+classical_svm = QuantumClassifierWithDefaultRiemannianPipeline(
+    shots=None,  # 'None' forces classic SVM
+    dim_red=PCA(n_components=5),
+)
+
+# Piepeline 3
+vqc = QuantumClassifierWithDefaultRiemannianPipeline(
+    dim_red=PCA(n_components=5),
+    # These parameters are specific to VQC.
+    # The pipeline will detect this and instantiate a VQC under the hood
+    spsa_trials=40,
+    two_local_reps=3,
+)
+
+# Piepeline 4
+quantum_mdm = QuantumMDMWithRiemannianPipeline()
+
+# Piepeline 5
+mdm = make_pipeline(ERPCovariances(estimator="lwf"), MDM())
+
+classifiers = [vqc, quantum_svm, classical_svm, quantum_mdm, mdm]
+
+n_classifiers = len(classifiers)
+
+# https://stackoverflow.com/questions/61825227/plotting-multiple-confusion-matrix-side-by-side
+f, axes = plt.subplots(1, n_classifiers, sharey="row")
+
+disp = None
+
+# Compute results
+for idx in range(n_classifiers):
+    # Training and classification
+    clf = classifiers[idx]
+    clf.fit(X_train, y_train)
+    y_pred = clf.predict(X_test)
+
+    # Printing the results
+    acc = balanced_accuracy_score(y_test, y_pred)
+    acc_str = "%0.2f" % acc
+
+    # Results visualization
+    # A confusion matrix is reported for each classifier. A perfectly performing
+    # classifier will have only its diagonal filled and the rest will be zeros.
+    names = ["aud left", "aud right", "vis left", "vis right"]
+    title = (
+        ("VQC (" if idx == 0 else "Quantum SVM (" if idx == 1 else "Classical SVM (")
+        if idx == 2
+        else "Quantum MDM ("
+        if idx == 3
+        else "R-MDM (" + acc_str + ")"
+    )
+    axe = axes[idx]
+    cm = confusion_matrix(y_pred, y_test)
+    disp = ConfusionMatrixDisplay(cm, display_labels=names)
+    disp.plot(ax=axe, xticks_rotation=45)
+    disp.ax_.set_title(title)
+    disp.im_.colorbar.remove()
+    disp.ax_.set_xlabel("")
+    if idx > 0:
+        disp.ax_.set_ylabel("")
+
+# Display all the confusion matrices
+if disp:
+    f.text(0.4, 0.1, "Predicted label", ha="left")
+    plt.subplots_adjust(wspace=0.40, hspace=0.1)
+    f.colorbar(disp.im_, ax=axes)
+    plt.show()

examples/ERP/plot_classify_EEG_quantum_svm.py

@@ -6,7 +6,7 @@
 Decoding applied to EEG data in sensor space using RG.
 Xdawn spatial filtering is applied on covariances matrices, which are
 then projected in the tangent space and classified with a quantum SVM
- classifier. It is compared to the classical SVM on binary classification.
+classifier. It is compared to the classical SVM on binary classification.
 
 """
 # Author: Gregoire Cattan
@@ -67,7 +67,7 @@
     y_pred = clf.predict(X_test)
 
     # Printing the results
-    acc = balanced_accuracy_score(y_pred, y_test)
+    acc = balanced_accuracy_score(y_test, y_pred)
     acc_str = "%0.2f" % acc
 
     names = ["vis left", "vis right"]

pyriemann_qiskit/classification.py

@@ -1,4 +1,9 @@
-"""Module for classification function."""
+"""
+Contains the base class for all quantum classifiers
+as well as several quantum classifiers than can run
+in several modes quantum/classical and simulated/real
+quantum computer.
+"""
 import numpy as np
 from sklearn.base import BaseEstimator, ClassifierMixin
 from sklearn.svm import SVC
@@ -37,13 +42,10 @@ class QuanticClassifierBase(BaseEstimator, ClassifierMixin):
     * tasks on a real quantum computer are assigned to a queue
       before being executed on a back-end (delayed execution)
 
-    WARNING: At the moment this implementation only supports binary
-    classification.
-
     Parameters
     ----------
     quantum : bool (default: True)
-        - If true will run on local or remote backend
+        - If true will run on local or remote quantum backend
         (depending on q_account_token value).
         - If false, will perform classical computing instead
     q_account_token : string (default:None)
@@ -62,6 +64,8 @@ class QuanticClassifierBase(BaseEstimator, ClassifierMixin):
     Notes
     -----
     .. versionadded:: 0.0.1
+    .. versionchanged:: 0.1.0
+        Added support for multi-class classification.
 
     Attributes
     ----------
@@ -120,24 +124,24 @@ def _log(self, *values):
             print("[QClass] ", *values)
 
     def _split_classes(self, X, y):
-        self._log(
-            "[Warning] Splitting first class from second class."
-            "Only binary classification is supported."
-        )
-        X_class1 = X[y == self.classes_[1]]
-        X_class0 = X[y == self.classes_[0]]
-        return (X_class1, X_class0)
+        n_classes = len(self.classes_)
+        X_classes = []
+        for idx in range(n_classes):
+            X_classes.append(X[y == self.classes_[idx]])
+        return X_classes
 
-    def _map_classes_to_0_1(self, y):
+    def _map_classes_to_indices(self, y):
         y_copy = y.copy()
-        y_copy[y == self.classes_[0]] = 0
-        y_copy[y == self.classes_[1]] = 1
+        n_classes = len(self.classes_)
+        for idx in range(n_classes):
+            y_copy[y == self.classes_[idx]] = idx
         return y_copy
 
-    def _map_0_1_to_classes(self, y):
+    def _map_indices_to_classes(self, y):
         y_copy = y.copy()
-        y_copy[y == 0] = self.classes_[0]
-        y_copy[y == 1] = self.classes_[1]
+        n_classes = len(self.classes_)
+        for idx in range(n_classes):
+            y_copy[y == idx] = self.classes_[idx]
         return y_copy
 
     def fit(self, X, y):
@@ -151,11 +155,6 @@ def fit(self, X, y):
         y : ndarray, shape (n_samples,)
             Target vector relative to X.
 
-        Raises
-        ------
-        Exception
-            Raised if the number of classes is different from 2
-
         Returns
         -------
         self : QuanticClassifierBase instance
@@ -165,17 +164,13 @@ def fit(self, X, y):
 
         self._log("Fitting: ", X.shape)
         self.classes_ = np.unique(y)
-        if len(self.classes_) != 2:
-            raise Exception(
-                "Only binary classification \
-                             is currently supported."
-            )
 
-        class1, class0 = self._split_classes(X, y)
-        y = self._map_classes_to_0_1(y)
+        X_classes = self._split_classes(X, y)
+        y = self._map_classes_to_indices(y)
 
-        self._training_input[self.classes_[1]] = class1
-        self._training_input[self.classes_[0]] = class0
+        n_classes = len(self.classes_)
+        for idx in range(n_classes):
+            self._training_input[self.classes_[idx]] = X_classes[idx]
 
         n_features = get_feature_dimension(self._training_input)
         self._log("Feature dimension = ", n_features)
@@ -226,7 +221,7 @@ def score(self, X, y):
         accuracy : double
             Accuracy of predictions from X with respect y.
         """
-        y = self._map_classes_to_0_1(y)
+        y = self._map_classes_to_indices(y)
         self._log("Testing...")
         return self._classifier.score(X, y)
 
@@ -399,7 +394,7 @@ def predict(self, X):
             Class labels for samples in X.
         """
         labels = self._predict(X)
-        return self._map_0_1_to_classes(labels)
+        return self._map_indices_to_classes(labels)
 
 
 class QuanticVQC(QuanticClassifierBase):
@@ -439,6 +434,7 @@ class QuanticVQC(QuanticClassifierBase):
     .. versionadded:: 0.0.1
     .. versionchanged:: 0.1.0
        Fix: copy estimator not keeping base class parameters.
+       Added support for multi-class classification.
 
     See Also
     --------
@@ -498,20 +494,6 @@ def _init_algo(self, n_features):
         )
         return vqc
 
-    def _map_classes_to_0_1(self, y):
-        # Label must be one-hot encoded for VQC
-        y_copy = np.ndarray((y.shape[0], 2))
-        y_copy[y == self.classes_[0]] = [1, 0]
-        y_copy[y == self.classes_[1]] = [0, 1]
-        return y_copy
-
-    def _map_0_1_to_classes(self, y):
-        # Decode one-hot encoded labels
-        y_copy = np.ndarray((y.shape[0], 1))
-        y_copy[(y == [1, 0]).all()] = self.classes_[0]
-        y_copy[(y == [0, 1]).all()] = self.classes_[1]
-        return y_copy
-
     def predict_proba(self, X):
         """Returns the probabilities associated with predictions.
 
@@ -545,7 +527,7 @@ def predict(self, X):
             Class labels for samples in X.
         """
         labels = self._predict(X)
-        return self._map_0_1_to_classes(labels)
+        return self._map_indices_to_classes(labels)
 
 
 class QuanticMDM(QuanticClassifierBase):
@@ -670,4 +652,4 @@ def predict(self, X):
             Class labels for samples in X.
         """
         labels = self._predict(X)
-        return self._map_0_1_to_classes(labels)
+        return self._map_indices_to_classes(labels)

pyriemann_qiskit/datasets/utils.py

@@ -1,3 +1,6 @@
+"""
+Contains helper methods and classes to manage datasets.
+"""
 from warnings import warn
 import numpy as np
 
@@ -10,7 +13,7 @@
 from sklearn.datasets import make_classification
 
 
-def get_mne_sample(n_trials=10):
+def get_mne_sample(n_trials=10, include_auditory=False):
     """Return sample data from the mne dataset.
 
     ```
@@ -31,6 +34,9 @@ def get_mne_sample(n_trials=10):
     n_trials : int (default:10)
         Number of trials to return.
         If -1, then all trials are returned.
+    include_auditory : boolean (default:False)
+        If True, it returns also the auditory stimulation
+        in the MNE dataset.
 
     Returns
     -------
@@ -42,6 +48,8 @@ def get_mne_sample(n_trials=10):
     Notes
     -----
     .. versionadded:: 0.0.1
+    .. versionchanged:: 0.1.0
+        Possibility to include auditory stimulation.
 
     References
     ----------
@@ -55,7 +63,10 @@ def get_mne_sample(n_trials=10):
     raw_fname = data_path + "/MEG/sample/sample_audvis_filt-0-40_raw.fif"
     event_fname = data_path + "/MEG/sample/sample_audvis_filt-0-40_raw-eve.fif"
     tmin, tmax = -0.0, 1
-    event_id = dict(vis_l=3, vis_r=4)  # select only two classes
+    if include_auditory:
+        event_id = dict(aud_l=1, aud_r=2, vis_l=3, vis_r=4)
+    else:
+        event_id = dict(vis_l=3, vis_r=4)
 
     # Setup for reading the raw data
     raw = io.Raw(raw_fname, preload=True, verbose=False)

pyriemann_qiskit/pipelines.py

@@ -185,10 +185,14 @@ class QuantumClassifierWithDefaultRiemannianPipeline(BasePipeline):
     feature_reps : int (default: 2)
         The number of repeated circuits for the ZZFeatureMap,
         greater or equal to 1.
-    spsa_trials : int (default: 40)
+    spsa_trials : int (default: None)
         Maximum number of iterations to perform using SPSA optimizer.
-    two_local_reps : int (default: 3)
+        For VQC, you can use 40 as a default.
+        VQC is only enabled if spsa_trials and two_local_reps are not None.
+    two_local_reps : int (default: None)
         The number of repetition for the two-local cricuit.
+        VQC is only enabled if spsa_trials and two_local_reps are not None.
+        For VQC, you can use 3 as a default.
     params: Dict (default: {})
         Additional parameters to pass to the nested instance
         of the quantum classifier.