08-08 细分构建机器学习应用程序的流程-模型优化

细分构建机器学习应用程序的流程-模型优化

  通过数据收集、数据预处理、训练模型、测试模型上述四个步骤，一般可以得到一个不错的模型，但是一般得到的都是一个参数收敛的模型，然而我们模型还有超参数或不同的核函数等，如r的非线性支持向量机的bf核或linear核；rbf核的非线性支持向量机超参数$C、\gamma$，正则化中的$\alpha$。我们模型优化一块主要是对模型超参数的优化，简而言之就是输入一组超参数，对每个超参数对应的模型进行测试，选择这一组超参数中最优的模型。

1.1 网格搜索法

  网格搜索法相当于对你输入的每一个参数都进行验证，并且可以设置多个超参数。

from sklearn import datasets
from sklearn.svm import SVC
from sklearn.model_selection import ShuffleSplit
from sklearn.model_selection import GridSearchCV

iris = datasets.load_iris()
X = iris.data
y = iris.target

# 总共有2*4=8种选择
parameters = {'kernel': ('linear', 'rbf'), 'C': [0.1, 1, 10, 100]}

svc = SVC(gamma="scale")

cv = ShuffleSplit(n_splits=10, test_size=0.3, random_state=1)

scoring = 'accuracy'

clf = GridSearchCV(svc, parameters, cv=cv, scoring=scoring)
clf.fit(X, y)

GridSearchCV(cv=ShuffleSplit(n_splits=10, random_state=1, test_size=0.3, train_size=None),
       error_score='raise-deprecating',
       estimator=SVC(C=1.0, cache_size=200, class_weight=None, coef0=0.0,
  decision_function_shape='ovr', degree=3, gamma='scale', kernel='rbf',
  max_iter=-1, probability=False, random_state=None, shrinking=True,
  tol=0.001, verbose=False),
       fit_params=None, iid='warn', n_jobs=None,
       param_grid={'kernel': ('linear', 'rbf'), 'C': [0.1, 1, 10, 100]},
       pre_dispatch='2*n_jobs', refit=True, return_train_score='warn',
       scoring='accuracy', verbose=0)

clf.cv_results_.keys()

dict_keys(['mean_fit_time', 'std_fit_time', 'mean_score_time', 'std_score_time', 'param_kernel', 'param_gamma', 'param_C', 'params', 'split0_test_score', 'split1_test_score', 'split2_test_score', 'split3_test_score', 'split4_test_score', 'mean_test_score', 'std_test_score', 'rank_test_score', 'split0_train_score', 'split1_train_score', 'split2_train_score', 'split3_train_score', 'split4_train_score', 'mean_train_score', 'std_train_score'])

clf.best_params_

{'C': 1, 'kernel': 'linear'}

clf.best_score_

0.9844444444444445

1.2 随机搜索法

  随机搜索法一般用于超参数过多的时候，即当一组超参数有上千上万个的时候，我们会通过随机搜索法随机选择一部分超参数，对模型调优。一般随机搜索法会通过sklearn.model_selection.ParameterSampler方法进行采样。

1.2.1 随机采样

from sklearn.model_selection import ParameterSampler
from scipy.stats.distributions import expon
import numpy as np

np.random.seed(1)
param_grid = {'a': [1, 2], 'b': expon()}
# expon为指数分布，该分布取值为无数个，即param_grid有无数个参数
param_grid

{'a': [1, 2],
 'b': <scipy.stats._distn_infrastructure.rv_frozen at 0x1a1772a630>}

# n_iter=4指定采样次数为4次
param_list = list(ParameterSampler(param_grid, n_iter=5))

rounded_list = [dict((k, round(v, 2)) for (k, v) in d.items())
                for d in param_list]
rounded_list

[{'a': 2, 'b': 5.87},
 {'a': 1, 'b': 0.0},
 {'a': 2, 'b': 6.95},
 {'a': 1, 'b': 0.1},
 {'a': 1, 'b': 0.49}]

1.2.2 随机搜索法

from sklearn import datasets
from sklearn.svm import SVC
from sklearn.model_selection import ShuffleSplit
from sklearn.model_selection import RandomizedSearchCV

iris = datasets.load_iris()
X = iris.data
y = iris.target

# 总共有2*5*5=50种选择
parameters = {'kernel': ('linear', 'rbf'), 'C': [
    0.1, 1, 10, 100, 1000], 'gamma': [0.1, 1, 10, 100, 1000]}
svc = SVC()

cv = ShuffleSplit(n_splits=5, test_size=0.3, random_state=1)

scoring = 'accuracy'

clf = RandomizedSearchCV(svc, parameters, cv=cv, scoring=scoring, n_iter=15)
clf.fit(X, y)

RandomizedSearchCV(cv=ShuffleSplit(n_splits=5, random_state=1, test_size=0.3, train_size=None),
          error_score='raise-deprecating',
          estimator=SVC(C=1.0, cache_size=200, class_weight=None, coef0=0.0,
  decision_function_shape='ovr', degree=3, gamma='auto_deprecated',
  kernel='rbf', max_iter=-1, probability=False, random_state=None,
  shrinking=True, tol=0.001, verbose=False),
          fit_params=None, iid='warn', n_iter=15, n_jobs=None,
          param_distributions={'kernel': ('linear', 'rbf'), 'C': [0.1, 1, 10, 100, 1000], 'gamma': [0.1, 1, 10, 100, 1000]},
          pre_dispatch='2*n_jobs', random_state=None, refit=True,
          return_train_score='warn', scoring='accuracy', verbose=0)

clf.cv_results_.keys()

dict_keys(['mean_fit_time', 'std_fit_time', 'mean_score_time', 'std_score_time', 'param_kernel', 'param_gamma', 'param_C', 'params', 'split0_test_score', 'split1_test_score', 'split2_test_score', 'split3_test_score', 'split4_test_score', 'mean_test_score', 'std_test_score', 'rank_test_score', 'split0_train_score', 'split1_train_score', 'split2_train_score', 'split3_train_score', 'split4_train_score', 'mean_train_score', 'std_train_score'])

clf.best_params_

{'kernel': 'linear', 'gamma': 1000, 'C': 1}

clf.best_score_

0.9822222222222222