RBF 核的顯式特徵圖近似#

一個說明 RBF 核的特徵圖近似的範例。

它展示如何使用 RBFSampler 和 Nystroem 來近似 RBF 核的特徵圖，以使用 SVM 對數字資料集進行分類。比較了原始空間中線性 SVM、使用近似映射的線性 SVM 和使用核化 SVM 的結果。顯示了使用不同數量的蒙特卡羅採樣（在 RBFSampler 的情況下，它使用隨機傅立葉特徵）和近似映射的不同大小訓練集子集（對於 Nystroem）的計時和準確性。

請注意，此處的資料集不夠大，無法顯示核近似的好處，因為精確的 SVM 仍然相當快。

採樣更多維度顯然會帶來更好的分類結果，但成本更高。這意味著運行時間和準確性之間存在權衡，由參數 n_components 給定。請注意，使用 SGDClassifier 通過隨機梯度下降可以大大加速線性 SVM 和近似核 SVM 的求解。對於核化 SVM 的情況，這並不容易實現。

Python 套件和資料集匯入、載入資料集#

# Authors: The scikit-learn developers
# SPDX-License-Identifier: BSD-3-Clause

# Standard scientific Python imports
from time import time

import matplotlib.pyplot as plt
import numpy as np

# Import datasets, classifiers and performance metrics
from sklearn import datasets, pipeline, svm
from sklearn.decomposition import PCA
from sklearn.kernel_approximation import Nystroem, RBFSampler

# The digits dataset
digits = datasets.load_digits(n_class=9)

計時和準確性圖#

為了在此資料上應用分類器，我們需要將影像展平，將資料轉換為 (樣本，特徵) 矩陣

n_samples = len(digits.data)
data = digits.data / 16.0
data -= data.mean(axis=0)

# We learn the digits on the first half of the digits
data_train, targets_train = (data[: n_samples // 2], digits.target[: n_samples // 2])


# Now predict the value of the digit on the second half:
data_test, targets_test = (data[n_samples // 2 :], digits.target[n_samples // 2 :])
# data_test = scaler.transform(data_test)

# Create a classifier: a support vector classifier
kernel_svm = svm.SVC(gamma=0.2)
linear_svm = svm.LinearSVC(random_state=42)

# create pipeline from kernel approximation
# and linear svm
feature_map_fourier = RBFSampler(gamma=0.2, random_state=1)
feature_map_nystroem = Nystroem(gamma=0.2, random_state=1)
fourier_approx_svm = pipeline.Pipeline(
    [
        ("feature_map", feature_map_fourier),
        ("svm", svm.LinearSVC(random_state=42)),
    ]
)

nystroem_approx_svm = pipeline.Pipeline(
    [
        ("feature_map", feature_map_nystroem),
        ("svm", svm.LinearSVC(random_state=42)),
    ]
)

# fit and predict using linear and kernel svm:

kernel_svm_time = time()
kernel_svm.fit(data_train, targets_train)
kernel_svm_score = kernel_svm.score(data_test, targets_test)
kernel_svm_time = time() - kernel_svm_time

linear_svm_time = time()
linear_svm.fit(data_train, targets_train)
linear_svm_score = linear_svm.score(data_test, targets_test)
linear_svm_time = time() - linear_svm_time

sample_sizes = 30 * np.arange(1, 10)
fourier_scores = []
nystroem_scores = []
fourier_times = []
nystroem_times = []

for D in sample_sizes:
    fourier_approx_svm.set_params(feature_map__n_components=D)
    nystroem_approx_svm.set_params(feature_map__n_components=D)
    start = time()
    nystroem_approx_svm.fit(data_train, targets_train)
    nystroem_times.append(time() - start)

    start = time()
    fourier_approx_svm.fit(data_train, targets_train)
    fourier_times.append(time() - start)

    fourier_score = fourier_approx_svm.score(data_test, targets_test)
    nystroem_score = nystroem_approx_svm.score(data_test, targets_test)
    nystroem_scores.append(nystroem_score)
    fourier_scores.append(fourier_score)

# plot the results:
plt.figure(figsize=(16, 4))
accuracy = plt.subplot(121)
# second y axis for timings
timescale = plt.subplot(122)

accuracy.plot(sample_sizes, nystroem_scores, label="Nystroem approx. kernel")
timescale.plot(sample_sizes, nystroem_times, "--", label="Nystroem approx. kernel")

accuracy.plot(sample_sizes, fourier_scores, label="Fourier approx. kernel")
timescale.plot(sample_sizes, fourier_times, "--", label="Fourier approx. kernel")

# horizontal lines for exact rbf and linear kernels:
accuracy.plot(
    [sample_sizes[0], sample_sizes[-1]],
    [linear_svm_score, linear_svm_score],
    label="linear svm",
)
timescale.plot(
    [sample_sizes[0], sample_sizes[-1]],
    [linear_svm_time, linear_svm_time],
    "--",
    label="linear svm",
)

accuracy.plot(
    [sample_sizes[0], sample_sizes[-1]],
    [kernel_svm_score, kernel_svm_score],
    label="rbf svm",
)
timescale.plot(
    [sample_sizes[0], sample_sizes[-1]],
    [kernel_svm_time, kernel_svm_time],
    "--",
    label="rbf svm",
)

# vertical line for dataset dimensionality = 64
accuracy.plot([64, 64], [0.7, 1], label="n_features")

# legends and labels
accuracy.set_title("Classification accuracy")
timescale.set_title("Training times")
accuracy.set_xlim(sample_sizes[0], sample_sizes[-1])
accuracy.set_xticks(())
accuracy.set_ylim(np.min(fourier_scores), 1)
timescale.set_xlabel("Sampling steps = transformed feature dimension")
accuracy.set_ylabel("Classification accuracy")
timescale.set_ylabel("Training time in seconds")
accuracy.legend(loc="best")
timescale.legend(loc="best")
plt.tight_layout()
plt.show()

RBF 核 SVM 和線性 SVM 的決策面#

第二個圖視覺化了 RBF 核 SVM 和具有近似核圖的線性 SVM 的決策面。該圖顯示了分類器投影到資料的前兩個主成分上的決策面。此視覺化應謹慎看待，因為它只是通過 64 個維度的決策面的有趣切片。特別要注意的是，資料點（表示為點）不一定會被分類到它所在的區域，因為它不會位於前兩個主成分跨越的平面上。在核近似中詳細描述了 RBFSampler 和 Nystroem 的用法。

# visualize the decision surface, projected down to the first
# two principal components of the dataset
pca = PCA(n_components=8, random_state=42).fit(data_train)

X = pca.transform(data_train)

# Generate grid along first two principal components
multiples = np.arange(-2, 2, 0.1)
# steps along first component
first = multiples[:, np.newaxis] * pca.components_[0, :]
# steps along second component
second = multiples[:, np.newaxis] * pca.components_[1, :]
# combine
grid = first[np.newaxis, :, :] + second[:, np.newaxis, :]
flat_grid = grid.reshape(-1, data.shape[1])

# title for the plots
titles = [
    "SVC with rbf kernel",
    "SVC (linear kernel)\n with Fourier rbf feature map\nn_components=100",
    "SVC (linear kernel)\n with Nystroem rbf feature map\nn_components=100",
]

plt.figure(figsize=(18, 7.5))
plt.rcParams.update({"font.size": 14})
# predict and plot
for i, clf in enumerate((kernel_svm, nystroem_approx_svm, fourier_approx_svm)):
    # Plot the decision boundary. For that, we will assign a color to each
    # point in the mesh [x_min, x_max]x[y_min, y_max].
    plt.subplot(1, 3, i + 1)
    Z = clf.predict(flat_grid)

    # Put the result into a color plot
    Z = Z.reshape(grid.shape[:-1])
    levels = np.arange(10)
    lv_eps = 0.01  # Adjust a mapping from calculated contour levels to color.
    plt.contourf(
        multiples,
        multiples,
        Z,
        levels=levels - lv_eps,
        cmap=plt.cm.tab10,
        vmin=0,
        vmax=10,
        alpha=0.7,
    )
    plt.axis("off")

    # Plot also the training points
    plt.scatter(
        X[:, 0],
        X[:, 1],
        c=targets_train,
        cmap=plt.cm.tab10,
        edgecolors=(0, 0, 0),
        vmin=0,
        vmax=10,
    )

    plt.title(titles[i])
plt.tight_layout()
plt.show()