Fix/svc kernel (#220)

* - Use x-feature-map - Change predict to use predict_proba instead for QSVM * - test for predict proba - add multiclass FVT for QSVM * - fix predict_proba for multiclass - use predict_proba inside for qsvm predictions - add multiclass test for MDM * fix conftest * fix MDM multiclass test * reactivate quantum pipeline with plot_classify_P300_bi.py * fix labels for multiclass * [pre-commit.ci] auto fixes from pre-commit.com hooks * remove feature_entanglement from plot_compare_dim * missing cast to np.array * [pre-commit.ci] auto fixes from pre-commit.com hooks * - fix financial example (inverted classifier rf and erp) - only apply softmax on predict (remove useless overriden method predict_proba inside QSVM). * fix _map_indices_to_classes * [pre-commit.ci] auto fixes from pre-commit.com hooks * Anton's suggestion * [pre-commit.ci] auto fixes from pre-commit.com hooks * various typo * - Run multiclass classification example: Ok - fix titles * [pre-commit.ci] auto fixes from pre-commit.com hooks * minor correction --------- Co-authored-by: Gregoire Cattan <gregoire.cattan@ibm.com> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: qbarthelemy <q.barthelemy@gmail.com>
pyRiemann · Dec 8, 2023 · 7deee15 · 7deee15
1 parent cfeb7bd
commit 7deee15
Show file tree

Hide file tree

Showing 9 changed files with 133 additions and 117 deletions.
diff --git a/README.md b/README.md
@@ -96,7 +96,8 @@ This library is part of the [Qiskit Ecosystem](https://qiskit.org/ecosystem)
 _We recommend the use of [Anaconda](https://www.anaconda.com/) to manage python
 environements._
 
-`pyRiemann-qiskit` currently supports Windows, Mac and Linux OS with Python 3.9 - 3.11.
+`pyRiemann-qiskit` currently supports Windows, Mac and Linux OS with **Python 3.9 -
+3.11**.
 
 You can install `pyRiemann-qiskit` release from PyPI:
 
@@ -105,13 +106,8 @@ pip install pyriemann-qiskit
 ```
 
 The development version can be installed by cloning this repository and installing the
-package on your local machine using the `setup.py` script:
-
-```
-python setup.py develop
-```
-
-Or directly pip:
+package on your local machine using the `setup.py` script. We recommand to do it using
+`pip`:
 
 ```
 pip install .

diff --git a/examples/ERP/plot_classify_P300_bi.py b/examples/ERP/plot_classify_P300_bi.py
@@ -67,7 +67,7 @@
 
 # reduce the number of subjects, the Quantum pipeline takes a lot of time
 # if executed on the entire dataset
-n_subjects = 5
+n_subjects = 2
 for dataset in datasets:
     dataset.subject_list = dataset.subject_list[0:n_subjects]
 
@@ -78,7 +78,7 @@
 # A Riemannian Quantum pipeline provided by pyRiemann-qiskit
 # You can choose between classical SVM and Quantum SVM.
 pipelines["RG+QuantumSVM"] = QuantumClassifierWithDefaultRiemannianPipeline(
-    shots=None,  # 'None' forces classic SVM
+    shots=512,  # 'None' forces classic SVM
     nfilter=2,  # default 2
     # default n_components=10, a higher value renders better performance with
     # the non-qunatum SVM version used in qiskit

diff --git a/examples/MI/multiclass_classification.py b/examples/MI/multiclass_classification.py
@@ -99,13 +99,21 @@
     # 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 + ")"
-    )
+    if idx == 0:
+        title = "VQC"
+    elif idx == 1:
+        title = "Q-SVM"
+    elif idx == 2:
+        title = "SVM"
+    elif idx == 3:
+        title = "Q-MDM"
+    else:
+        title = "MDM"
+
+    title = f"{title} (" + acc_str + ")"
+
+    print(title)
+
     axe = axes[idx]
     cm = confusion_matrix(y_pred, y_test)
     disp = ConfusionMatrixDisplay(cm, display_labels=names)

diff --git a/examples/MI/plot_compare_dim_red.py b/examples/MI/plot_compare_dim_red.py
@@ -39,11 +39,7 @@
     # Determine the number of "run" on the quantum machine (simulated or real)
     # the higher is this number, the lower the variability.
     "shots": [1024],  # [512, 1024, 2048]
-    # This defines how we entangle the data into a quantum state
-    # the more complex is the kernel, the more outcomes we can expect from
-    # a quantum vs classical classifier.
-    "feature_entanglement": ["linear"],  # ['linear', 'sca', 'full'],
-    # This parameter change the depth of the circuit when entangling data.
+    # This parameter changes the depth of the circuit when entangling data.
     # There is a trade-off between accuracy and noise when the depth of the
     # circuit increases.
     "feature_reps": [2],  # [2, 3, 4]

diff --git a/examples/other_datasets/plot_financial_data.py b/examples/other_datasets/plot_financial_data.py
@@ -443,7 +443,7 @@ def transform(self, X):
 
 
 class ERP_CollusionClassifier(ClassifierMixin):
-    def __init__(self, row_clf, erp_clf, threshold=0.5):
+    def __init__(self, erp_clf, row_clf, threshold=0.5):
         self.row_clf = row_clf
         self.erp_clf = erp_clf
         self.threshold = threshold
@@ -453,7 +453,7 @@ def fit(self, X, y):
         return self
 
     def predict(self, X):
-        y_pred = self.row_clf.predict(X)
+        y_pred = self.row_clf.predict(X).astype(float)
         collusion_prob = self.erp_clf.predict_proba(X)
         y_pred[y_pred == 1] = collusion_prob[y_pred == 1, 1].transpose()
         y_pred[y_pred >= self.threshold] = 1

diff --git a/pyriemann_qiskit/classification.py b/pyriemann_qiskit/classification.py
@@ -5,6 +5,7 @@
 quantum computer.
 """
 from datetime import datetime
+from scipy.special import softmax
 import logging
 import numpy as np
 
@@ -145,11 +146,11 @@ def _map_classes_to_indices(self, y):
         return y_copy
 
     def _map_indices_to_classes(self, y):
-        y_copy = y.copy()
+        y_copy = np.array(y.copy())
         n_classes = len(self.classes_)
         for idx in range(n_classes):
-            y_copy[y == idx] = self.classes_[idx]
-        return y_copy
+            y_copy[np.array(y).transpose() == idx] = self.classes_[idx]
+        return np.array(y_copy)
 
     def fit(self, X, y):
         """Uses a quantum backend and fits the training data.
@@ -237,6 +238,49 @@ def _predict(self, X):
         self._log("Prediction finished.")
         return result
 
+    def predict_proba(self, X):
+        """Return the probabilities associated with predictions.
+
+        The default behavior is to return the nested classifier probabilities.
+        In case where no `predict_proba` method is available inside the classifier,
+        the method predicts the label number (0 or 1 for examples) and applies a
+        softmax in top of it.
+
+        Parameters
+        ----------
+        X : ndarray, shape (n_samples, n_features)
+            Input vector, where `n_samples` is the number of samples and
+            `n_features` is the number of features.
+
+        Returns
+        -------
+        prob : ndarray, shape (n_samples, n_classes)
+            prob[n, i] == 1 if the nth sample is assigned to class `i`.
+        """
+
+        if not hasattr(self._classifier, "predict_proba"):
+            # Classifier has no predict_proba
+            # Use the result from predict and apply a softmax
+            self._log(
+                "No predict_proba method available.\
+                       Computing softmax probabilities..."
+            )
+            proba = self._classifier.predict(X)
+            proba = [
+                np.array(
+                    [
+                        1 if c == self.classes_[i] else 0
+                        for i in range(len(self.classes_))
+                    ]
+                )
+                for c in proba
+            ]
+            proba = softmax(proba, axis=0)
+        else:
+            proba = self._classifier.predict_proba(X)
+
+        return np.array(proba)
+
 
 class QuanticSVM(QuanticClassifierBase):
 
@@ -256,6 +300,8 @@ class QuanticSVM(QuanticClassifierBase):
         Fix: copy estimator not keeping base class parameters.
     .. versionchanged:: 0.2.0
         Add seed parameter
+        SVC and QSVC now compute probability (may impact performance)
+        Predict is now using predict_proba with a softmax, when using QSVC.
 
     Parameters
     ----------
@@ -360,39 +406,15 @@ def _init_algo(self, n_features):
                     gamma=self.gamma,
                     C=self.C,
                     max_iter=max_iter,
+                    probability=True,
                 )
         else:
             max_iter = -1 if self.max_iter is None else self.max_iter
-            classifier = SVC(gamma=self.gamma, C=self.C, max_iter=max_iter)
+            classifier = SVC(
+                gamma=self.gamma, C=self.C, max_iter=max_iter, probability=True
+            )
         return classifier
 
-    def predict_proba(self, X):
-        """Return the probabilities associated with predictions.
-
-        This method is implemented for compatibility purpose
-        as SVM prediction probabilities are not available.
-        This method assigns a boolean value to each trial which
-        depends on whether the label was assigned to class 0 or 1
-
-        Parameters
-        ----------
-        X : ndarray, shape (n_samples, n_features)
-            Input vector, where `n_samples` is the number of samples and
-            `n_features` is the number of features.
-
-        Returns
-        -------
-        prob : ndarray, shape (n_samples, n_classes)
-            prob[n, 0] == True if the nth sample is assigned to 1st class;
-            prob[n, 1] == True if the nth sample is assigned to 2nd class.
-        """
-        predicted_labels = self.predict(X)
-        ret = [
-            np.array([c == self.classes_[0], c == self.classes_[1]])
-            for c in predicted_labels
-        ]
-        return np.array(ret)
-
     def predict(self, X):
         """Calculates the predictions.
 
@@ -407,7 +429,12 @@ def predict(self, X):
         pred : array, shape (n_samples,)
             Class labels for samples in X.
         """
-        labels = self._predict(X)
+        if isinstance(self._classifier, QSVC):
+            probs = softmax(self.predict_proba(X))
+            labels = [np.argmax(prob) for prob in probs]
+        else:
+            labels = self._predict(X)
+        self._log("Prediction finished.")
         return self._map_indices_to_classes(labels)
 
 
@@ -514,24 +541,6 @@ def _init_algo(self, n_features):
         )
         return vqc
 
-    def predict_proba(self, X):
-        """Returns the probabilities associated with predictions.
-
-        Parameters
-        ----------
-        X : ndarray, shape (n_samples, n_features)
-            Input vector, where `n_samples` is the number of samples and
-            `n_features` is the number of features.
-
-        Returns
-        -------
-        prob : ndarray, shape (n_samples, n_classes)
-            prob[n, 0] == True if the nth sample is assigned to 1st class;
-            prob[n, 1] == True if the nth sample is assigned to 2nd class.
-        """
-        proba, _ = self._predict(X)
-        return proba
-
     def predict(self, X):
         """Calculates the predictions.
 
@@ -664,22 +673,6 @@ def _init_algo(self, n_features):
         set_global_optimizer(self._optimizer)
         return classifier
 
-    def predict_proba(self, X):
-        """Return the probabilities associated with predictions.
-
-        Parameters
-        ----------
-        X : ndarray, shape (n_trials, n_channels, n_channels)
-            ndarray of trials.
-
-        Returns
-        -------
-        prob : ndarray, shape (n_samples, n_classes)
-            prob[n, 0] == True if the nth sample is assigned to 1st class;
-            prob[n, 1] == True if the nth sample is assigned to 2nd class.
-        """
-        return self._classifier.predict_proba(X)
-
     def predict(self, X):
         """Calculates the predictions.
 

diff --git a/pyriemann_qiskit/pipelines.py b/pyriemann_qiskit/pipelines.py
@@ -10,6 +10,7 @@
 from pyriemann_qiskit.utils.filtering import NoDimRed
 from pyriemann_qiskit.utils.hyper_params_factory import (
     gen_zz_feature_map,
+    gen_x_feature_map,
     gen_two_local,
     get_spsa,
 )
@@ -175,14 +176,8 @@ class QuantumClassifierWithDefaultRiemannianPipeline(BasePipeline):
     shots : int | None (default: 1024)
         Number of repetitions of each circuit, for sampling.
         If None, classical computation will be performed.
-    feature_entanglement : str | list[list[list[int]]] | \
-                   Callable[int, list[list[list[int]]]]
-        Specifies the entanglement structure for the ZZFeatureMap.
-        Entanglement structure can be provided with indices or string.
-        Possible string values are: 'full', 'linear', 'circular' and 'sca'.
-        See [2]_ for more details on entanglement structure.
     feature_reps : int (default: 2)
-        The number of repeated circuits for the ZZFeatureMap,
+        The number of repeated circuits for the FeatureMap,
         greater or equal to 1.
     spsa_trials : int (default: None)
         Maximum number of iterations to perform using SPSA optimizer.
@@ -206,12 +201,15 @@ class QuantumClassifierWithDefaultRiemannianPipeline(BasePipeline):
     Notes
     -----
     .. versionadded:: 0.0.1
+    .. versionchanged:: 0.2.0
+        Changed feature map from ZZFeatureMap to XFeatureMap.
+        Therefore remove unused parameter `feature_entanglement`.
 
     See Also
     --------
     XdawnCovariances
     TangentSpace
-    gen_zz_feature_map
+    gen_x_feature_map
     gen_two_local
     get_spsa
     QuanticVQC
@@ -236,7 +234,6 @@ def __init__(
         C=1.0,
         max_iter=None,
         shots=1024,
-        feature_entanglement="full",
         feature_reps=2,
         spsa_trials=None,
         two_local_reps=None,
@@ -248,7 +245,6 @@ def __init__(
         self.C = C
         self.max_iter = max_iter
         self.shots = shots
-        self.feature_entanglement = feature_entanglement
         self.feature_reps = feature_reps
         self.spsa_trials = spsa_trials
         self.two_local_reps = two_local_reps
@@ -261,7 +257,11 @@ def _create_pipe(self):
         is_vqc = self.spsa_trials and self.two_local_reps
         is_quantum = self.shots is not None
 
-        feature_map = gen_zz_feature_map(self.feature_reps, self.feature_entanglement)
+        # Different feature maps can be used.
+        # Currently the best results are produced by the x_feature_map.
+        # This can change in the future as the code for the different feature maps
+        # is updated in the new versions of Qiskit.
+        feature_map = gen_x_feature_map(self.feature_reps)
 
         if is_vqc:
             self._log("QuanticVQC chosen.")
@@ -320,7 +320,7 @@ class QuantumMDMWithRiemannianPipeline(BasePipeline):
     shots : int (default:1024)
         Number of repetitions of each circuit, for sampling.
     gen_feature_map : Callable[int, QuantumCircuit | FeatureMap] \
-                      (default : Callable[int, ZZFeatureMap])
+                      (default : Callable[int, XFeatureMap])
         Function generating a feature map to encode data into a quantum state.
 
     Attributes

diff --git a/tests/conftest.py b/tests/conftest.py
@@ -100,8 +100,9 @@ def _get_dataset(n_samples, n_features, n_classes, type="bin"):
             samples_0 = make_covariances(
                 n_samples // n_classes, n_features, 0, return_params=False
             )
-            samples_1 = samples_0 * 2
-            samples = np.concatenate((samples_0, samples_1), axis=0)
+            samples = np.concatenate(
+                [samples_0 * (i + 1) for i in range(n_classes)], axis=0
+            )
             labels = _get_labels(n_samples, n_classes)
         else:
             samples, labels = get_mne_sample()
@@ -179,6 +180,7 @@ class BinaryFVT(BinaryTest):
     def additional_steps(self):
         self.quantum_instance.fit(self.samples, self.labels)
         self.prediction = self.quantum_instance.predict(self.samples)
+        self.predict_proab = self.quantum_instance.predict_proba(self.samples)
         print(self.labels, self.prediction)