sklearn实现lda模型_使用python+sklearn实现模型复杂性的影响

本示例演示了模型复杂性是如何影响预测精度和计算性能的。 数据集是用于回归(或者分类)的波士顿住房(Boston Housing)数据集(或者20 Newsgroups)。 对于每一类模型，我们通过选择相关的模型参数来改变模型的复杂度，并测量对计算性能(延迟)和预测能力(MSE或Hamming损失)的影响。

输出：

Benchmarking SGDClassifier(alpha=0.001, l1_ratio=0.25, loss='modified_huber',
              penalty='elasticnet')
Complexity: 4466 | Hamming Loss (Misclassification Ratio): 0.2491 | Pred. Time: 0.020835s
Benchmarking SGDClassifier(alpha=0.001, l1_ratio=0.5, loss='modified_huber',
              penalty='elasticnet')
Complexity: 1663 | Hamming Loss (Misclassification Ratio): 0.2915 | Pred. Time: 0.015789s
Benchmarking SGDClassifier(alpha=0.001, l1_ratio=0.75, loss='modified_huber',
              penalty='elasticnet')
Complexity: 880 | Hamming Loss (Misclassification Ratio): 0.3180 | Pred. Time: 0.013469s
Benchmarking SGDClassifier(alpha=0.001, l1_ratio=0.9, loss='modified_huber',
              penalty='elasticnet')
Complexity: 639 | Hamming Loss (Misclassification Ratio): 0.3337 | Pred. Time: 0.011812s
Benchmarking NuSVR(C=1000.0, gamma=3.0517578125e-05, nu=0.1)
Complexity: 69 | MSE: 31.8139 | Pred. Time: 0.000301s
Benchmarking NuSVR(C=1000.0, gamma=3.0517578125e-05, nu=0.25)
Complexity: 136 | MSE: 25.6140 | Pred. Time: 0.000811s
Benchmarking NuSVR(C=1000.0, gamma=3.0517578125e-05)
Complexity: 244 | MSE: 22.3375 | Pred. Time: 0.000895s
Benchmarking NuSVR(C=1000.0, gamma=3.0517578125e-05, nu=0.75)
Complexity: 351 | MSE: 21.3688 | Pred. Time: 0.001237s
Benchmarking NuSVR(C=1000.0, gamma=3.0517578125e-05, nu=0.9)
Complexity: 404 | MSE: 21.1033 | Pred. Time: 0.001460s
Benchmarking GradientBoostingRegressor(n_estimators=10)
Complexity: 10 | MSE: 29.0148 | Pred. Time: 0.000120s
Benchmarking GradientBoostingRegressor(n_estimators=50)
Complexity: 50 | MSE: 8.6545 | Pred. Time: 0.000302s
Benchmarking GradientBoostingRegressor()
Complexity: 100 | MSE: 7.7179 | Pred. Time: 0.000264s
Benchmarking GradientBoostingRegressor(n_estimators=200)
Complexity: 200 | MSE: 6.7507 | Pred. Time: 0.000425s
Benchmarking GradientBoostingRegressor(n_estimators=500)
Complexity: 500 | MSE: 7.1471 | Pred. Time: 0.000922s

print(__doc__)# 作者: Eustache Diemert # 许可证: BSD 3 clause
import time
import numpy as np
import matplotlib.pyplot as plt
from mpl_toolkits.axes_grid1.parasite_axes import host_subplot
from mpl_toolkits.axisartist.axislines import Axes
from scipy.sparse.csr import csr_matrix
from sklearn import datasets
from sklearn.utils import shuffle
from sklearn.metrics import mean_squared_error
from sklearn.svm import NuSVR
from sklearn.ensemble import GradientBoostingRegressor
from sklearn.linear_model import SGDClassifier
from sklearn.metrics import hamming_loss# ############################################################################## 例程(Routines)# 初始化随机生成器
np.random.seed(0)
def generate_data(case, sparse=False):"""生成回归/分类数据。"""if case == 'regression':
        X, y = datasets.load_boston(return_X_y=True)elif case == 'classification':
        X, y = datasets.fetch_20newsgroups_vectorized(subset='all',
                                                      return_X_y=True)
    X, y = shuffle(X, y)
    offset = int(X.shape[0] * 0.8)
    X_train, y_train = X[:offset], y[:offset]
    X_test, y_test = X[offset:], y[offset:]if sparse:
        X_train = csr_matrix(X_train)
        X_test = csr_matrix(X_test)else:
        X_train = np.array(X_train)
        X_test = np.array(X_test)
    y_test = np.array(y_test)
    y_train = np.array(y_train)
    data = {'X_train': X_train, 'X_test': X_test, 'y_train': y_train,'y_test': y_test}return data
def benchmark_influence(conf):"""
    changing_param：对MSE和延迟的基准影响
    """
    prediction_times = []
    prediction_powers = []
    complexities = []for param_value in conf['changing_param_values']:
        conf['tuned_params'][conf['changing_param']] = param_value
        estimator = conf['estimator'](**conf['tuned_params'])print("Benchmarking %s" % estimator)
        estimator.fit(conf['data']['X_train'], conf['data']['y_train'])
        conf['postfit_hook'](estimator)
        complexity = conf['complexity_computer'](estimator)
        complexities.append(complexity)
        start_time = time.time()for _ in range(conf['n_samples']):
            y_pred = estimator.predict(conf['data']['X_test'])
        elapsed_time = (time.time() - start_time) / float(conf['n_samples'])
        prediction_times.append(elapsed_time)
        pred_score = conf['prediction_performance_computer'](
            conf['data']['y_test'], y_pred)
        prediction_powers.append(pred_score)print("Complexity: %d | %s: %.4f | Pred. Time: %fs\n" % (
            complexity, conf['prediction_performance_label'], pred_score,
            elapsed_time))return prediction_powers, prediction_times, complexities
def plot_influence(conf, mse_values, prediction_times, complexities):"""
    绘制模型复杂性对准确度和延迟的影响。
    """
    plt.figure(figsize=(12, 6))
    host = host_subplot(111, axes_class=Axes)
    plt.subplots_adjust(right=0.75)
    par1 = host.twinx()
    host.set_xlabel('Model Complexity (%s)' % conf['complexity_label'])
    y1_label = conf['prediction_performance_label']
    y2_label = "Time (s)"
    host.set_ylabel(y1_label)
    par1.set_ylabel(y2_label)
    p1, = host.plot(complexities, mse_values, 'b-', label="prediction error")
    p2, = par1.plot(complexities, prediction_times, 'r-',
                    label="latency")
    host.legend(loc='upper right')
    host.axis["left"].label.set_color(p1.get_color())
    par1.axis["right"].label.set_color(p2.get_color())
    plt.title('Influence of Model Complexity - %s' % conf['estimator'].__name__)
    plt.show()
def _count_nonzero_coefficients(estimator):
    a = estimator.coef_.toarray()return np.count_nonzero(a)# ############################################################################## 主要代码
regression_data = generate_data('regression')
classification_data = generate_data('classification', sparse=True)
configurations = [
    {'estimator': SGDClassifier,'tuned_params': {'penalty': 'elasticnet', 'alpha': 0.001, 'loss':'modified_huber', 'fit_intercept': True, 'tol': 1e-3},'changing_param': 'l1_ratio','changing_param_values': [0.25, 0.5, 0.75, 0.9],'complexity_label': 'non_zero coefficients','complexity_computer': _count_nonzero_coefficients,'prediction_performance_computer': hamming_loss,'prediction_performance_label': 'Hamming Loss (Misclassification Ratio)','postfit_hook': lambda x: x.sparsify(),'data': classification_data,'n_samples': 30},
    {'estimator': NuSVR,'tuned_params': {'C': 1e3, 'gamma': 2 ** -15},'changing_param': 'nu','changing_param_values': [0.1, 0.25, 0.5, 0.75, 0.9],'complexity_label': 'n_support_vectors','complexity_computer': lambda x: len(x.support_vectors_),'data': regression_data,'postfit_hook': lambda x: x,'prediction_performance_computer': mean_squared_error,'prediction_performance_label': 'MSE','n_samples': 30},
    {'estimator': GradientBoostingRegressor,'tuned_params': {'loss': 'ls'},'changing_param': 'n_estimators','changing_param_values': [10, 50, 100, 200, 500],'complexity_label': 'n_trees','complexity_computer': lambda x: x.n_estimators,'data': regression_data,'postfit_hook': lambda x: x,'prediction_performance_computer': mean_squared_error,'prediction_performance_label': 'MSE','n_samples': 30},
]for conf in configurations:
    prediction_performances, prediction_times, complexities = \
        benchmark_influence(conf)
    plot_influence(conf, prediction_performances, prediction_times,
                   complexities)

脚本的总运行时间： ( 0 分 20.261 秒) 估计的内存使用量： 60 MB 下载python源代码:plot_model_complexity_influence.py 下载Jupyter notebook源代码:plot_model_complexity_influence.ipynb 由Sphinx-Gallery生成的画廊 ☆☆☆为方便大家查阅，小编已将scikit-learn学习路线专栏文章统一整理到公众号底部菜单栏，同步更新中，关注公众号，点击左下方“系列文章”，如图： 欢迎大家和我一起沿着scikit-learn文档这条路线，一起巩固机器学习算法基础。(添加微信：mthler，备注：sklearn学习，一起进【sklearn机器学习进步群】开启打怪升级的学习之旅。) ‍