詳解Python中4種超參自動優化演演算法的實現

from sklearn import svm, datasets
from sklearn.model_selection import GridSearchCV
import pandas as pd

# 匯入資料
iris = datasets.load_iris()
# 定義超參搜尋空間
parameters = {'kernel':('linear', 'rbf'), 'C':[1, 10]}
# 初始化模型
svc = svm.SVC()
# 網格搜尋
clf = GridSearchCV(estimator = svc,
                   param_grid = parameters,
                   scoring = 'accuracy',
                   n_jobs = -1,
                   cv = 5)
clf.fit(iris.data, iris.target)
返回：GridSearchCV(cv=5, estimator=SVC(), n_jobs=-1,
                   param_grid={'C': [1, 10], 'kernel': ('linear', 'rbf')},
                   scoring='accuracy')

# 列印結果
print('詳細結果:n', pd.DataFrame.from_dict(clf.cv_results_))
print('最佳分類器:n', clf.best_estimator_)
print('最佳分數:n', clf.best_score_)
print('最佳引數:n', clf.best_params_).
返回：
詳細結果:
    mean_fit_time  std_fit_time  mean_score_time  std_score_time param_C  ... split3_test_score split4_test_score  mean_test_score  std_test_score  rank_test_score
0       0.000788      0.000394         0.000194        0.000389       1  ...          0.966667               1.0         0.980000        0.016330                1
1       0.000804      0.000402         0.000199        0.000399       1  ...          0.933333               1.0         0.966667        0.021082                4
2       0.000593      0.000484         0.000593        0.000484      10  ...          0.966667               1.0         0.973333        0.038873                3
3       0.000593      0.000484         0.000399        0.000489      10  ...          0.966667               1.0         0.980000        0.016330                1
[4 rows x 15 columns]
最佳分類器:
 SVC(C=1, kernel='linear')
最佳分數:
0.9800000000000001
最佳引數:
 {'C': 1, 'kernel': 'linear'}

from sklearn import svm, datasets
from sklearn.model_selection import RandomizedSearchCV
import pandas as pd
from scipy.stats import uniform


# 匯入資料
iris = datasets.load_iris()
# 定義超參搜尋空間
distributions = {'kernel':['linear', 'rbf'], 'C':uniform(loc=1, scale=9)}
# 初始化模型
svc = svm.SVC()
# 網格搜尋
clf = RandomizedSearchCV(estimator = svc,
                         param_distributions = distributions,
                         n_iter = 4,
                         scoring = 'accuracy',
                         cv = 5,
                         n_jobs = -1,
                         random_state = 2021)
clf.fit(iris.data, iris.target)
返回：RandomizedSearchCV(cv=5, estimator=SVC(), n_iter=4, n_jobs=-1,
                         param_distributions={'C': <scipy.stats._distn_infrastructure.rv_frozen object at 0x000001F372F9A190>,
                                              'kernel': ['linear', 'rbf']},
                         random_state=2021, scoring='accuracy')


# 列印結果
print('詳細結果:n', pd.DataFrame.from_dict(clf.cv_results_))
print('最佳分類器:n', clf.best_estimator_)
print('最佳分數:n', clf.best_score_)
print('最佳引數:n', clf.best_params_)
返回：
詳細結果:
    mean_fit_time  std_fit_time  mean_score_time  std_score_time  param_C  ... split3_test_score split4_test_score  mean_test_score  std_test_score  rank_test_score
0       0.000598      0.000489         0.000200        0.000400   6.4538  ...          0.966667               1.0         0.986667        0.016330                1
1       0.000997      0.000002         0.000000        0.000000  4.99782  ...          0.966667               1.0         0.980000        0.026667                3
2       0.000798      0.000399         0.000399        0.000488  3.81406  ...          0.966667               1.0         0.980000        0.016330                3
3       0.000598      0.000488         0.000200        0.000399  5.36286  ...          0.966667               1.0         0.986667        0.016330                1
[4 rows x 15 columns]
最佳分類器:
 SVC(C=6.453804509266643)
最佳分數:
0.9866666666666667
最佳引數:
 {'C': 6.453804509266643, 'kernel': 'rbf'}

from sklearn import svm, datasets
from sklearn.model_selection import cross_val_score
from hyperopt import hp, fmin, tpe, space_eval
import pandas as pd

# 匯入資料
iris = datasets.load_iris()

# step1: 定義目標函數
def objective(params):
      # 初始化模型並交叉驗證
      svc = svm.SVC(**params)
      cv_scores = cross_val_score(svc, iris.data, iris.target, cv=5)
      # 返回loss = 1 - accuracy (loss必須被最小化)
      loss = 1 - cv_scores.mean()
      return loss

# step2: 定義超參搜尋空間
space = {'kernel':hp.choice('kernel', ['linear', 'rbf']),
         'C':hp.uniform('C', 1, 10)}

# step3: 在給定超參搜尋空間下，最小化目標函數
best = fmin(objective, space, algo=tpe.suggest, max_evals=100)
返回: best_loss: 0.013333333333333308(即accuracy為0.9866666666666667)

# step4: 列印結果
print(best)
返回:{'C': 6.136181888987526, 'kernel': 1}(PS:kernel為0-based index,這裡1指rbf)