网格搜索和逐次减半的比较#

本示例比较了 HalvingGridSearchCV 和 GridSearchCV 执行的参数搜索。

from time import time

import matplotlib.pyplot as plt
import numpy as np
import pandas as pd

from sklearn import datasets
from sklearn.experimental import enable_halving_search_cv  # noqa
from sklearn.model_selection import GridSearchCV, HalvingGridSearchCV
from sklearn.svm import SVC

我们首先定义 SVC 估计器的参数空间，并计算训练 HalvingGridSearchCV 实例和 GridSearchCV 实例所需的时间。

rng = np.random.RandomState(0)
X, y = datasets.make_classification(n_samples=1000, random_state=rng)

gammas = [1e-1, 1e-2, 1e-3, 1e-4, 1e-5, 1e-6, 1e-7]
Cs = [1, 10, 100, 1e3, 1e4, 1e5]
param_grid = {"gamma": gammas, "C": Cs}

clf = SVC(random_state=rng)

tic = time()
gsh = HalvingGridSearchCV(
    estimator=clf, param_grid=param_grid, factor=2, random_state=rng
)
gsh.fit(X, y)
gsh_time = time() - tic

tic = time()
gs = GridSearchCV(estimator=clf, param_grid=param_grid)
gs.fit(X, y)
gs_time = time() - tic

现在我们为两个搜索估计器绘制热图。

def make_heatmap(ax, gs, is_sh=False, make_cbar=False):
    """Helper to make a heatmap."""
    results = pd.DataFrame(gs.cv_results_)
    results[["param_C", "param_gamma"]] = results[["param_C", "param_gamma"]].astype(
        np.float64
    )
    if is_sh:
        # SH dataframe: get mean_test_score values for the highest iter
        scores_matrix = results.sort_values("iter").pivot_table(
            index="param_gamma",
            columns="param_C",
            values="mean_test_score",
            aggfunc="last",
        )
    else:
        scores_matrix = results.pivot(
            index="param_gamma", columns="param_C", values="mean_test_score"
        )

    im = ax.imshow(scores_matrix)

    ax.set_xticks(np.arange(len(Cs)))
    ax.set_xticklabels(["{:.0E}".format(x) for x in Cs])
    ax.set_xlabel("C", fontsize=15)

    ax.set_yticks(np.arange(len(gammas)))
    ax.set_yticklabels(["{:.0E}".format(x) for x in gammas])
    ax.set_ylabel("gamma", fontsize=15)

    # Rotate the tick labels and set their alignment.
    plt.setp(ax.get_xticklabels(), rotation=45, ha="right", rotation_mode="anchor")

    if is_sh:
        iterations = results.pivot_table(
            index="param_gamma", columns="param_C", values="iter", aggfunc="max"
        ).values
        for i in range(len(gammas)):
            for j in range(len(Cs)):
                ax.text(
                    j,
                    i,
                    iterations[i, j],
                    ha="center",
                    va="center",
                    color="w",
                    fontsize=20,
                )

    if make_cbar:
        fig.subplots_adjust(right=0.8)
        cbar_ax = fig.add_axes([0.85, 0.15, 0.05, 0.7])
        fig.colorbar(im, cax=cbar_ax)
        cbar_ax.set_ylabel("mean_test_score", rotation=-90, va="bottom", fontsize=15)


fig, axes = plt.subplots(ncols=2, sharey=True)
ax1, ax2 = axes

make_heatmap(ax1, gsh, is_sh=True)
make_heatmap(ax2, gs, make_cbar=True)

ax1.set_title("Successive Halving\ntime = {:.3f}s".format(gsh_time), fontsize=15)
ax2.set_title("GridSearch\ntime = {:.3f}s".format(gs_time), fontsize=15)

plt.show()

Successive Halving time = 1.560s, GridSearch time = 6.653s

热图显示了 SVC 实例的参数组合的平均测试分数。 HalvingGridSearchCV 还显示了最后使用组合的迭代次数。标记为 0 的组合仅在第一次迭代中进行评估，而标记为 5 的组合被认为是最佳参数组合。

我们可以看到，HalvingGridSearchCV 类能够在更短的时间内找到与 GridSearchCV 一样准确的参数组合。

脚本的总运行时间：（0 分钟 8.400 秒）