网格搜索[6]_sklearn

网格搜索[6]_sklearn

1. 什么是网格搜索？

• Grid Search：一种调参手段；穷举搜索：在所有候选的参数选择中，通过循环遍历，尝试每一种可能性，表现最好的参数就是最终的结果。其原理就像是在数组里找最大值。（为什么叫网格搜索？以有两个参数的模型为例，参数a有3种可能，参数b有4种可能，把所有可能性列出来，可以表示成一个3*4的表格，其中每个cell就是一个网格，循环过程就像是在每个网格里遍历、搜索，所以叫grid search）

import pandas as pd
from sklearn.model_selection import train_test_split
from sklearn.svm import SVC
from sklearn.preprocessing import StandardScaler
'''------------------------------------------
1 读取数据
----------------------------------------------'''
data = pd.read_csv('data_processed.csv',encoding='gbk') 
'''-------------------------------------------
1.1 划分训练集何验证集
----------------------------------------------'''
train, test = train_test_split(data, test_size=0.3, random_state=666)
'''----------------------------------------
1.2 获取标签
-------------------------------------------'''
y_train = train.status
train.drop(['status'], axis=1, inplace=True)
y_test = test.status
test.drop(['status'], axis=1, inplace=True)
'''---------------------------------------------
1.3 数据标准化
-----------------------------------------------'''
scaler = StandardScaler()
train = pd.DataFrame(scaler.fit_transform(train),index=train.index, columns=test.columns)
test = pd.DataFrame(scaler.fit_transform(test),index=test.index, columns=test.columns)
'''----------------------------------------
1.4 训练模型
------------------------------------------'''
best_score = 0
for gamma in [0.001,0.01,0.1,1,10,100]:
    for C in [0.001,0.01,0.1,1,10,100]:
        model = SVC(gamma=gamma,C=C)
        model.fit(train, y_train)
        score = model.score(test, y_test)
        if score > best_score:  #找到最好表现的参数
            best_score = score
            best_parameters = {'gamma':gamma,'c':C}
'''--------------------------------------
1.5 预测模型
---------------------------------------------'''
y_test_pre = model.predict(test)
'''------------------------------------------
1.6 评分
---------------------------------------------'''
print('-------------------------------')
print('best score验证分数：{}'.format(best_score))
print("best parameters:{}".format(best_parameters))

1.1存在问题：

• 原始数据集划分训练集和测试集以后，其中测试集除了用作参数调整，也用来测试模型的好坏，这样做导致最终的评分结果比实际的效果要好

解决办法：

• 对训练集再进行一次划分，分成训练集和验证集：
  • 训练集用来训练模型，验证集用来调整参数，测试集用来衡量模型好坏
    

import pandas as pd
from sklearn.model_selection import train_test_split
from sklearn.svm import SVC
from sklearn.preprocessing import StandardScaler
'''------------------------------------------
1 读取数据
----------------------------------------------'''
data = pd.read_csv('data_processed.csv',encoding='gbk') 
'''-------------------------------------------
1.1 划分训练集、验证集、测试集
----------------------------------------------'''
train_val, test = train_test_split(data, test_size=0.1, random_state=666)
train, val = train_test_split(train_val,test_size=0.1, random_state=666)
'''----------------------------------------
1.2 获取标签
-------------------------------------------'''
y_train_val = train_val.status                  # 训练集+验证集
train_val.drop(['status'], axis=1,inplace=True)
y_train = train.status
train.drop(['status'], axis=1, inplace=True)
y_test = test.status
test.drop(['status'], axis=1, inplace=True)
y_val = val.status
val.drop(['status'], axis=1, inplace = True)
'''---------------------------------------------
1.3 数据标准化
-----------------------------------------------'''
scaler = StandardScaler()
train = pd.DataFrame(scaler.fit_transform(train),index=train.index, columns=test.columns)
test = pd.DataFrame(scaler.fit_transform(test),index=test.index, columns=test.columns)
train_val = pd.DataFrame(scaler.fit_transform(train_val),index=train_val.index, columns=train_val.columns)
val = pd.DataFrame(scaler.fit_transform(val),index=val.index, columns=val.columns)
'''----------------------------------------
1.4 训练模型
------------------------------------------'''
best_score = 0
for gamma in [0.001,0.01,0.1,1,10,100]:
    for C in [0.001,0.01,0.1,1,10,100]:
        model = SVC(gamma=gamma,C=C)
        model.fit(train, y_train)
        score = model.score(test, y_test)
        if score > best_score:  #找到最好表现的参数
            best_score = score
            best_parameters = {'gamma':gamma,'C':C}
'''--------------------------------------
1.5 预测模型
---------------------------------------------'''
model = SVC(**best_parameters)  #使用最好参数，构建新的模型
model.fit(train_val, y_train_val)    #使用训练集和验证集进行训练
best_score = model.score(test,y_test)
print('-------------------------------')
print('best score验证分数：{}'.format(best_score))
print("best parameters:{}".format(best_parameters))

• 注意数据的标准化，标准化后，数据准确率显著提升

2. 交叉验证

• 交叉验证法：

  1. 将数据集D划分为k个大小相似的互斥子集，即$D=D_1\cup D_2\cup ...\cup D_k，D_i\cap D_j=\varnothing (i̸=j)$
  2. 每次用k-1个子集的并集作为训练集，余下的那个作为测试集
     • 这样就可以获得k组训练/测试集——最终返回的是这k个测试集结果的均值
     • 通常把交叉验证法称为“K折交叉验证”
       

• 留一法：

  • 假设数据集D中包含m个样本，若领k=m
  • 留一法不受随机样本划分方式的影响

• Grid Search with cross validation

from sklearn.model_selection import cross_val_score
'''-----------------------------------------
1.4 训练模型
------------------------------------------'''
best_score = 0
for gamma in [0.001,0.01,0.1,1,10,100]:
    for C in [0.001,0.01,0.1,1,10,100]:
        model = SVC(gamma=gamma,C=C)
        scores = cross_val_score(model, train, y_train, cv=5)
        score = scores.mean()
        if score > best_score:  #找到最好表现的参数
            best_score = score
            best_parameters = {'gamma':gamma,'C':C}
'''--------------------------------------
1.5 预测模型
---------------------------------------------'''
mpdel = SVC(**best_parameters)
model.fit(train,y_train)
test_score = model.score(test,y_test)

参考

调参必备–Girdsearch网格搜索