Proposal : Quantum Autoencoder (#280)

* Feat/qae (#36) * add autoencoder based on qiskit-ml example: https://qiskit.org/ecosystem/machine-learning/tutorials/12_quantum_autoencoder.html turned to be compatible with scikit-learn * set up example with autoencoder. Works with qiskit ml 0.7.1 * add dataset from https://zenodo.org/records/5055046 (partially implemented) * change localisation of the dataset. adapt exapmle to use the ergonomic dataset * apply denoising on raw signal * reduce epoch dimension use cross session evaluation * update pipeline * update optimizer * add rs and fix some bugs * add epochselectchannel * doc * improve code for channel selection * expose qaoa optimizer * push example * [pre-commit.ci] auto fixes from pre-commit.com hooks * typo in example * fix Attributes section duplicated * use moabb Hinss2021 * autoencoder.py Co-authored-by: Yash Chauhan <116367139+jiggychauhi@users.noreply.github.com> Co-authored-by: Ahmed Mostafa <121709141+AhmedMostafa167@users.noreply.github.com> * refactor autoencoder, add documentation * Complete documentation * extract devectorizer to preprocessing.py WIP * documentation for devectorizer * complete doc for EpochSelectChannel * refactor qword example. Do not use xDawn. * fix bug with spatial filter * add callback parameter * [pre-commit.ci] auto fixes from pre-commit.com hooks * typo * flake8 * [pre-commit.ci] auto fixes from pre-commit.com hooks * typo * flake8 * [pre-commit.ci] auto fixes from pre-commit.com hooks --------- 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: Yash Chauhan <116367139+jiggychauhi@users.noreply.github.com> Co-authored-by: Ahmed Mostafa <121709141+AhmedMostafa167@users.noreply.github.com> * Update examples/ERP/autoencoders_example.py Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * Update examples/ERP/autoencoders_example.py Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * Update examples/ERP/autoencoders_example.py Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * Update pyriemann_qiskit/autoencoders.py Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * Update pyriemann_qiskit/utils/preprocessing.py Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * Update pyriemann_qiskit/utils/filtering.py Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * Update pyriemann_qiskit/autoencoders.py Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * Update pyriemann_qiskit/autoencoders.py Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * Update pyriemann_qiskit/autoencoders.py Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * Update pyriemann_qiskit/autoencoders.py Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * Update pyriemann_qiskit/autoencoders.py Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * Update pyriemann_qiskit/autoencoders.py Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * Update pyriemann_qiskit/utils/filtering.py Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * Update pyriemann_qiskit/utils/filtering.py Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * Rename to ChannelSelection * Update pyriemann_qiskit/utils/filtering.py Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * Update pyriemann_qiskit/utils/filtering.py Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * Update pyriemann_qiskit/utils/filtering.py Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * Update pyriemann_qiskit/utils/preprocessing.py Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * replace n_times -> n_samples * replace n_trials by n_matrices * apply review comment marked as outdated by GH * fix example * don't plot example as Hinss2021 is still on MOABB dev * remove unused alias for identitiy gate * Update pyriemann_qiskit/utils/filtering.py Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * Update pyriemann_qiskit/autoencoders.py Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * Update pyriemann_qiskit/autoencoders.py Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * Update pyriemann_qiskit/utils/filtering.py Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> --------- 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: Yash Chauhan <116367139+jiggychauhi@users.noreply.github.com> Co-authored-by: Ahmed Mostafa <121709141+AhmedMostafa167@users.noreply.github.com> Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com>
pyRiemann · May 24, 2024 · e595da4 · e595da4
1 parent e95e4bc
commit e595da4
Show file tree

Hide file tree

Showing 7 changed files with 617 additions and 35 deletions.
diff --git a/doc/api.rst b/doc/api.rst
@@ -21,6 +21,18 @@ Classification
     NearestConvexHull
 
 
+Autoencoders
+-------------
+.. _autoencoders_api:
+.. currentmodule:: pyriemann_qiskit.autoencoders
+
+.. autosummary::
+    :toctree: generated/
+    :template: class.rst
+
+    BasicQnnAutoencoder
+
+
 Pipelines
 ---------
 .. _pipelines_api:
@@ -88,6 +100,8 @@ Preprocessing
     :toctree: generated/
 
     NdRobustScaler
+    Vectorizer
+    Devectorizer
 
 Filtering
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -99,7 +113,7 @@ Filtering
 
     NoDimRed
     NaiveDimRed
-    Vectorizer
+    ChannelSelection
 
 Mean
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~

diff --git a/examples/ERP/noplot_autoencoders_example.py b/examples/ERP/noplot_autoencoders_example.py
@@ -0,0 +1,195 @@
+"""
+===============================================================================
+Quantum autoencoder for signal denoising
+===============================================================================
+This is a basic example of a quantum autoencoder for signal denoising,
+based on [1]_ and [2]_.
+
+There is no particular advantage in using QNN for such a task.
+
+This is experimental and should be used for research purpose only.
+
+"""
+
+# Authors: A. Mostafa, Y. Chauhan, W. Ahmed, and G. Cattan
+# License: BSD (3-clause)
+
+import logging
+import warnings
+
+from matplotlib import pyplot as plt
+from moabb import set_log_level
+from moabb.datasets import Hinss2021
+from moabb.evaluations import CrossSessionEvaluation
+from moabb.paradigms import RestingStateToP300Adapter
+from pyriemann.estimation import Covariances
+from pyriemann.tangentspace import TangentSpace
+from qiskit_algorithms.optimizers import COBYLA
+import seaborn as sns
+from sklearn.discriminant_analysis import LinearDiscriminantAnalysis as LDA
+from sklearn.pipeline import make_pipeline
+
+from pyriemann_qiskit.autoencoders import BasicQnnAutoencoder
+from pyriemann_qiskit.utils.filtering import ChannelSelection
+from pyriemann_qiskit.utils.preprocessing import Vectorizer, Devectorizer
+
+print(__doc__)
+
+##############################################################################
+# getting rid of the warnings about the future
+
+warnings.simplefilter(action="ignore", category=FutureWarning)
+warnings.simplefilter(action="ignore", category=RuntimeWarning)
+
+warnings.filterwarnings("ignore")
+
+set_log_level("info")
+
+##############################################################################
+# Setting logger level to info, to print the autoencoder trace.
+
+logging.getLogger("root").setLevel(logging.INFO)
+
+##############################################################################
+# Initialization
+# ----------------
+#
+# 1) Create paradigm
+# 2) Load datasets
+
+events = dict(easy=2, medium=3)
+paradigm = RestingStateToP300Adapter(events=events, tmin=0, tmax=0.5)
+
+datasets = [Hinss2021()]
+
+# reduce the number of subjects, the Quantum pipeline takes a lot of time
+# if executed on the entire dataset
+start_subject, stop_subject = 14, 15
+title = "Datasets: "
+for dataset in datasets:
+    title = title + " " + dataset.code
+    dataset.subject_list = dataset.subject_list[start_subject:stop_subject]
+
+##############################################################################
+# We have to do this because the classes are called 'Target' and 'NonTarget'
+# but the evaluation function uses a LabelEncoder, transforming them
+# to 0 and 1
+labels_dict = {"Target": 1, "NonTarget": 0}
+
+##############################################################################
+# Define a callback to keep trace of the computed costs.
+
+costs = {}
+
+
+def fn_callback(iter, cost):
+    if iter in costs:
+        costs[iter].append(cost)
+    else:
+        costs[iter] = [cost]
+
+
+##############################################################################
+# Create Pipelines
+# ----------------
+#
+# Pipelines must be a dict of sklearn pipeline transformer.
+
+pipelines = {}
+
+# An important limitation is that:
+# n_components x n_times = 2 ** (num_latent + num_trash)
+n_components, n_times = 8, 64
+
+pipelines["QNN+LDA"] = make_pipeline(
+    ChannelSelection(n_channels=n_components),
+    Vectorizer(),
+    # Use COBYLA with only 1 iteration (this is for runnin in Ci/Cd)
+    BasicQnnAutoencoder(
+        num_latent=n_components,
+        num_trash=1,
+        opt=COBYLA(maxiter=10),
+        callback=fn_callback,
+    ),
+    Devectorizer(n_components, n_times),
+    Covariances(),
+    TangentSpace(),
+    LDA(),
+)
+
+pipelines["LDA"] = make_pipeline(
+    ChannelSelection(n_channels=n_components),
+    Vectorizer(),
+    Devectorizer(n_components, n_times),
+    Covariances(),
+    TangentSpace(),
+    LDA(),
+)
+
+##############################################################################
+# Run evaluation
+# ----------------
+#
+# Compare the pipeline using a cross-sessions evaluation.
+
+evaluation = CrossSessionEvaluation(
+    paradigm=paradigm, datasets=datasets, overwrite=True, n_jobs=-1
+)
+
+results = evaluation.process(pipelines)
+
+print("Averaging the session performance:")
+print(results.groupby("pipeline").mean("score")[["score", "time"]])
+
+# ##############################################################################
+# Plot Results
+# ----------------
+#
+# Here we plot the results to compare two pipelines
+
+fig, ax = plt.subplots(facecolor="white", figsize=[8, 4])
+
+sns.stripplot(
+    data=results,
+    y="score",
+    x="pipeline",
+    ax=ax,
+    jitter=True,
+    alpha=0.5,
+    zorder=1,
+    palette="Set1",
+)
+sns.pointplot(data=results, y="score", x="pipeline", ax=ax, palette="Set1").set(
+    title=title
+)
+
+ax.set_ylabel("ROC AUC")
+ax.set_ylim(0.3, 1)
+
+plt.show()
+
+##############################################################################
+# Plot the mean cost function
+
+x = []
+y = []
+for iter in costs.keys():
+    x.append(iter)
+    c = costs[iter]
+    y.append(sum(c) / len(c))
+
+plt.plot(x, y)
+plt.xlabel("N of cost evaluation")
+plt.ylabel("Cost")
+plt.title("Autoencoder Cost")
+plt.tight_layout()
+plt.show()
+
+###############################################################################
+# References
+# ----------
+# .. [1] \
+#   https://qiskit-community.github.io/qiskit-machine-learning/tutorials/12_quantum_autoencoder.html
+# .. [2] A. Mostafa et al., 2024
+#   ‘Quantum Denoising in the Realm of Brain-Computer Interfaces: A Preliminary Study’,
+#   https://hal.science/hal-04501908
diff --git a/pyriemann_qiskit/__init__.py b/pyriemann_qiskit/__init__.py
@@ -1,4 +1,4 @@
 from ._version import __version__
-from . import classification, pipelines, ensemble
+from . import classification, pipelines, ensemble, autoencoders
 
-__all__ = ["__version__", "classification", "pipelines", "ensemble"]
+__all__ = ["__version__", "classification", "pipelines", "ensemble", "autoencoders"]