Nota

Esta página fue generada a partir de docs/tutorials/02_neural_network_classifier_and_regressor.ipynb.

Redes Neuronales de Clasificación y Regresión

En este tutorial mostramos cómo se utiliza la NeuralNetworkClassifier y la NeuralNetworkRegressor. Ambas toman como entrada una NeuralNetwork (Cuántica) y la aprovechan en un contexto específico. En ambos casos, también ofrecemos una variante preconfigurada para mayor comodidad, el Clasificador Cuántico Variacional (Variational Quantum Classifier, VQC) y el Regresor Cuántico Variacional (Variational Quantum Regressor, VQR). El tutorial está estructurado de la siguiente manera:

1. Clasificación

   • Classification with an EstimatorQNN

   • Classification with a SamplerQNN

   • Clasificador Cuántico Variacional (Variational Quantum Classifier, VQC)

2. Regresión

   • Regression with an EstimatorQNN

   • Regresor Cuántico Variacional (Variational Quantum Regressor, VQR)

import matplotlib.pyplot as plt
import numpy as np
from IPython.display import clear_output
from qiskit import QuantumCircuit
from qiskit.algorithms.optimizers import COBYLA, L_BFGS_B
from qiskit.circuit import Parameter
from qiskit.circuit.library import RealAmplitudes, ZZFeatureMap
from qiskit.utils import algorithm_globals

from qiskit_machine_learning.algorithms.classifiers import NeuralNetworkClassifier, VQC
from qiskit_machine_learning.algorithms.regressors import NeuralNetworkRegressor, VQR
from qiskit_machine_learning.neural_networks import SamplerQNN, EstimatorQNN

algorithm_globals.random_seed = 42

Clasificación

Preparamos un conjunto de datos de clasificación simple para ilustrar los siguientes algoritmos.

num_inputs = 2
num_samples = 20
X = 2 * algorithm_globals.random.random([num_samples, num_inputs]) - 1
y01 = 1 * (np.sum(X, axis=1) >= 0)  # in { 0,  1}
y = 2 * y01 - 1  # in {-1, +1}
y_one_hot = np.zeros((num_samples, 2))
for i in range(num_samples):
    y_one_hot[i, y01[i]] = 1

for x, y_target in zip(X, y):
    if y_target == 1:
        plt.plot(x[0], x[1], "bo")
    else:
        plt.plot(x[0], x[1], "go")
plt.plot([-1, 1], [1, -1], "--", color="black")
plt.show()

Classification with the an EstimatorQNN

First we show how an EstimatorQNN can be used for classification within a NeuralNetworkClassifier. In this context, the EstimatorQNN is expected to return one-dimensional output in \([-1, +1]\). This only works for binary classification and we assign the two classes to \(\{-1, +1\}\).

# construct QNN
qc = QuantumCircuit(2)
feature_map = ZZFeatureMap(2)
ansatz = RealAmplitudes(2)
qc.compose(feature_map, inplace=True)
qc.compose(ansatz, inplace=True)
qc.draw(output="mpl")

Create a quantum neural network

estimator_qnn = EstimatorQNN(
    circuit=qc, input_params=feature_map.parameters, weight_params=ansatz.parameters
)

# QNN maps inputs to [-1, +1]
estimator_qnn.forward(X[0, :], algorithm_globals.random.random(estimator_qnn.num_weights))

array([[0.23521988]])

Agregaremos una función de devolución de llamada nombrada callback_graph. Esto se llamará para cada iteración del optimizador y se le pasarán dos parámetros: los pesos actuales y el valor de la función objetivo en esos pesos. Para nuestra función, agregamos el valor de la función objetivo a un arreglo para que podamos graficar la iteración frente al valor de la función objetivo y actualizar la gráfica con cada iteración. Sin embargo, puedes hacer lo que desees con una función de devolución de llamada siempre y cuando se pasen los dos parámetros mencionados.

# callback function that draws a live plot when the .fit() method is called
def callback_graph(weights, obj_func_eval):
    clear_output(wait=True)
    objective_func_vals.append(obj_func_eval)
    plt.title("Objective function value against iteration")
    plt.xlabel("Iteration")
    plt.ylabel("Objective function value")
    plt.plot(range(len(objective_func_vals)), objective_func_vals)
    plt.show()

# construct neural network classifier
estimator_classifier = NeuralNetworkClassifier(
    estimator_qnn, optimizer=COBYLA(maxiter=60), callback=callback_graph
)

# create empty array for callback to store evaluations of the objective function
objective_func_vals = []
plt.rcParams["figure.figsize"] = (12, 6)

# fit classifier to data
estimator_classifier.fit(X, y)

# return to default figsize
plt.rcParams["figure.figsize"] = (6, 4)

# score classifier
estimator_classifier.score(X, y)

0.8

# evaluate data points
y_predict = estimator_classifier.predict(X)

# plot results
# red == wrongly classified
for x, y_target, y_p in zip(X, y, y_predict):
    if y_target == 1:
        plt.plot(x[0], x[1], "bo")
    else:
        plt.plot(x[0], x[1], "go")
    if y_target != y_p:
        plt.scatter(x[0], x[1], s=200, facecolors="none", edgecolors="r", linewidths=2)
plt.plot([-1, 1], [1, -1], "--", color="black")
plt.show()

Now, when the model is trained, we can explore the weights of the neural network. Please note, the number of weights is defined by ansatz.

estimator_classifier.weights

array([ 0.86209107, -1.06526254, -0.10663602, -0.39086371,  1.0894299 ,
        0.59368219,  2.22731471, -1.04769663])

Classification with a SamplerQNN

Next we show how a SamplerQNN can be used for classification within a NeuralNetworkClassifier. In this context, the SamplerQNN is expected to return \(d\)-dimensional probability vector as output, where \(d\) denotes the number of classes. The underlying Sampler primitive returns quasi-distributions of bit strings and we just need to define a mapping from the measured bitstrings to the different classes. For binary classification we use the parity mapping.

# construct feature map
feature_map = ZZFeatureMap(num_inputs)

# construct ansatz
ansatz = RealAmplitudes(num_inputs, reps=1)

# construct quantum circuit
qc = QuantumCircuit(num_inputs)
qc.append(feature_map, range(num_inputs))
qc.append(ansatz, range(num_inputs))
qc.decompose().draw(output="mpl")

# parity maps bitstrings to 0 or 1
def parity(x):
    return "{:b}".format(x).count("1") % 2


output_shape = 2  # corresponds to the number of classes, possible outcomes of the (parity) mapping.

# construct QNN
sampler_qnn = SamplerQNN(
    circuit=qc,
    input_params=feature_map.parameters,
    weight_params=ansatz.parameters,
    interpret=parity,
    output_shape=output_shape,
)

# construct classifier
sampler_classifier = NeuralNetworkClassifier(
    neural_network=sampler_qnn, optimizer=COBYLA(maxiter=30), callback=callback_graph
)

# create empty array for callback to store evaluations of the objective function
objective_func_vals = []
plt.rcParams["figure.figsize"] = (12, 6)

# fit classifier to data
sampler_classifier.fit(X, y01)

# return to default figsize
plt.rcParams["figure.figsize"] = (6, 4)

# score classifier
sampler_classifier.score(X, y01)

0.7

# evaluate data points
y_predict = sampler_classifier.predict(X)

# plot results
# red == wrongly classified
for x, y_target, y_p in zip(X, y01, y_predict):
    if y_target == 1:
        plt.plot(x[0], x[1], "bo")
    else:
        plt.plot(x[0], x[1], "go")
    if y_target != y_p:
        plt.scatter(x[0], x[1], s=200, facecolors="none", edgecolors="r", linewidths=2)
plt.plot([-1, 1], [1, -1], "--", color="black")
plt.show()

Again, once the model is trained we can take a look at the weights. As we set reps=1 explicitly in our ansatz, we can see less parameters than in the previous model.

sampler_classifier.weights

array([ 1.67198565,  0.46045402, -0.93462862, -0.95266092])

Clasificador Cuántico Variacional (Variational Quantum Classifier, VQC)

The VQC is a special variant of the NeuralNetworkClassifier with a SamplerQNN. It applies a parity mapping (or extensions to multiple classes) to map from the bitstring to the classification, which results in a probability vector, which is interpreted as a one-hot encoded result. By default, it applies this the CrossEntropyLoss function that expects labels given in one-hot encoded format and will return predictions in that format too.

# construct feature map, ansatz, and optimizer
feature_map = ZZFeatureMap(num_inputs)
ansatz = RealAmplitudes(num_inputs, reps=1)

# construct variational quantum classifier
vqc = VQC(
    feature_map=feature_map,
    ansatz=ansatz,
    loss="cross_entropy",
    optimizer=COBYLA(maxiter=30),
    callback=callback_graph,
)

# create empty array for callback to store evaluations of the objective function
objective_func_vals = []
plt.rcParams["figure.figsize"] = (12, 6)

# fit classifier to data
vqc.fit(X, y_one_hot)

# return to default figsize
plt.rcParams["figure.figsize"] = (6, 4)

# score classifier
vqc.score(X, y_one_hot)

0.8

# evaluate data points
y_predict = vqc.predict(X)

# plot results
# red == wrongly classified
for x, y_target, y_p in zip(X, y_one_hot, y_predict):
    if y_target[0] == 1:
        plt.plot(x[0], x[1], "bo")
    else:
        plt.plot(x[0], x[1], "go")
    if not np.all(y_target == y_p):
        plt.scatter(x[0], x[1], s=200, facecolors="none", edgecolors="r", linewidths=2)
plt.plot([-1, 1], [1, -1], "--", color="black")
plt.show()

Multiple classes with VQC

In this section we generate an artificial dataset that contains samples of three classes and show how to train a model to classify this dataset. This example shows how to tackle more interesting problems in machine learning. Of course, for a sake of short training time we prepare a tiny dataset. We employ make_classification from SciKit-Learn to generate a dataset. There 10 samples in the dataset, 2 features, that means we can still have a nice plot of the dataset, as well as no redundant features, these are features are generated as a combinations of the other features. Also, we have 3 different classes in the dataset, each classes one kind of centroid and we set class separation to 2.0, a slight increase from the default value of 1.0 to ease the classification problem.

Una vez que el conjunto de datos es generado, escalamos las características en el rango [0, 1].

from sklearn.datasets import make_classification
from sklearn.preprocessing import MinMaxScaler

X, y = make_classification(
    n_samples=10,
    n_features=2,
    n_classes=3,
    n_redundant=0,
    n_clusters_per_class=1,
    class_sep=2.0,
    random_state=algorithm_globals.random_seed,
)
X = MinMaxScaler().fit_transform(X)

Veamos cómo se ve nuestro conjunto de datos.

plt.scatter(X[:, 0], X[:, 1], c=y)

<matplotlib.collections.PathCollection at 0x24a97b228c8>

We also transform labels and make them categorical.

y_cat = np.empty(y.shape, dtype=str)
y_cat[y == 0] = "A"
y_cat[y == 1] = "B"
y_cat[y == 2] = "C"
print(y_cat)

['A' 'A' 'B' 'C' 'C' 'A' 'B' 'B' 'A' 'C']

We create an instance of VQC similar to the previous example, but in this case we pass a minimal set of parameters. Instead of feature map and ansatz we pass just the number of qubits that is equal to the number of features in the dataset, an optimizer with a low number of iteration to reduce training time, a quantum instance, and a callback to observe progress.

vqc = VQC(
    num_qubits=2,
    optimizer=COBYLA(maxiter=30),
    callback=callback_graph,
)

Start the training process in the same way as in previous examples.

# create empty array for callback to store evaluations of the objective function
objective_func_vals = []
plt.rcParams["figure.figsize"] = (12, 6)

# fit classifier to data
vqc.fit(X, y_cat)

# return to default figsize
plt.rcParams["figure.figsize"] = (6, 4)

# score classifier
vqc.score(X, y_cat)

0.9

A pesar de que tuvimos el bajo número de iteraciones, logramos una puntuación bastante buena. Veamos el resultado del método predict y comparemos el resultado con la verdad fundamental.

predict = vqc.predict(X)
print(f"Predicted labels: {predict}")
print(f"Ground truth:     {y_cat}")

Predicted labels: ['A' 'A' 'B' 'C' 'C' 'A' 'B' 'B' 'A' 'B']
Ground truth:     ['A' 'A' 'B' 'C' 'C' 'A' 'B' 'B' 'A' 'C']

Regresión

Preparamos un conjunto de datos de regresión simple para ilustrar los siguientes algoritmos.

num_samples = 20
eps = 0.2
lb, ub = -np.pi, np.pi
X_ = np.linspace(lb, ub, num=50).reshape(50, 1)
f = lambda x: np.sin(x)

X = (ub - lb) * algorithm_globals.random.random([num_samples, 1]) + lb
y = f(X[:, 0]) + eps * (2 * algorithm_globals.random.random(num_samples) - 1)

plt.plot(X_, f(X_), "r--")
plt.plot(X, y, "bo")
plt.show()

Regression with an EstimatorQNN

Here we restrict to regression with an EstimatorQNN that returns values in \([-1, +1]\). More complex and also multi-dimensional models could be constructed, also based on SamplerQNN but that exceeds the scope of this tutorial.

# construct simple feature map
param_x = Parameter("x")
feature_map = QuantumCircuit(1, name="fm")
feature_map.ry(param_x, 0)

# construct simple ansatz
param_y = Parameter("y")
ansatz = QuantumCircuit(1, name="vf")
ansatz.ry(param_y, 0)

# construct a circuit
qc = QuantumCircuit(1)
qc.compose(feature_map, inplace=True)
qc.compose(ansatz, inplace=True)

# construct QNN
regression_estimator_qnn = EstimatorQNN(
    circuit=qc, input_params=feature_map.parameters, weight_params=ansatz.parameters
)

# construct the regressor from the neural network
regressor = NeuralNetworkRegressor(
    neural_network=regression_estimator_qnn,
    loss="squared_error",
    optimizer=L_BFGS_B(maxiter=5),
    callback=callback_graph,
)

# create empty array for callback to store evaluations of the objective function
objective_func_vals = []
plt.rcParams["figure.figsize"] = (12, 6)

# fit to data
regressor.fit(X, y)

# return to default figsize
plt.rcParams["figure.figsize"] = (6, 4)

# score the result
regressor.score(X, y)

0.9769994291935521

# plot target function
plt.plot(X_, f(X_), "r--")

# plot data
plt.plot(X, y, "bo")

# plot fitted line
y_ = regressor.predict(X_)
plt.plot(X_, y_, "g-")
plt.show()

Similarly to the classification models, we can obtain an array of trained weights by querying a corresponding property of the model. In this model we have only one parameter defined as param_y above.

regressor.weights

array([-1.58870599])

Regresión con el Regresor Cuántico Variacional (Variational Quantum Regressor, VQR)

Similar to the VQC for classification, the VQR is a special variant of the NeuralNetworkRegressor with a EstimatorQNN. By default it considers the L2Loss function to minimize the mean squared error between predictions and targets.

vqr = VQR(
    feature_map=feature_map,
    ansatz=ansatz,
    optimizer=L_BFGS_B(maxiter=5),
    callback=callback_graph,
)

# create empty array for callback to store evaluations of the objective function
objective_func_vals = []
plt.rcParams["figure.figsize"] = (12, 6)

# fit regressor
vqr.fit(X, y)

# return to default figsize
plt.rcParams["figure.figsize"] = (6, 4)

# score result
vqr.score(X, y)

0.9769955693935385

# plot target function
plt.plot(X_, f(X_), "r--")

# plot data
plt.plot(X, y, "bo")

# plot fitted line
y_ = vqr.predict(X_)
plt.plot(X_, y_, "g-")
plt.show()

import qiskit.tools.jupyter

%qiskit_version_table
%qiskit_copyright

Version Information

Qiskit Software	Version
qiskit-terra	0.22.0
qiskit-aer	0.11.0
qiskit-ignis	0.7.0
qiskit	0.33.0
qiskit-machine-learning	0.5.0
System information
Python version	3.7.9
Python compiler	MSC v.1916 64 bit (AMD64)
Python build	default, Aug 31 2020 17:10:11
OS	Windows
CPUs	4
Memory (Gb)	31.837730407714844
Fri Oct 28 15:44:56 2022 GMT Daylight Time

This code is a part of Qiskit

© Copyright IBM 2017, 2022.

This code is licensed under the Apache License, Version 2.0. You may
obtain a copy of this license in the LICENSE.txt file in the root directory
of this source tree or at http://www.apache.org/licenses/LICENSE-2.0.

Any modifications or derivative works of this code must retain this
copyright notice, and modified files need to carry a notice indicating
that they have been altered from the originals.