机器学习4个常用超参数调试方法

ML工作流中最困难的部分之一是为模型找到最好的超参数。ML模型的性能与超参数直接相关。超参数调优的越好，得到的模型就越好。调优超参数可能是非常乏味和困难的，更像是一门艺术而不是科学。

超参数

超参数是在建立模型时用于控制算法行为的参数。这些参数不能从常规训练过程中获得。在对模型进行训练之前，需要对它们进行赋值。

1. 传统手工搜索

在传统的调参过程中，我们通过训练算法手动检查随机超参数集，并选择符合我们目标的最佳参数集。

我们看看代码：

#importing required libraries
from sklearn.neighbors import KNeighborsClassifier
from sklearn.model_selection import train_test_split
from sklearn.model_selection import KFold , cross_val_score
from sklearn.datasets import load_wine

wine = load_wine()
X = wine.data
y = wine.target

#splitting the data into train and test set
X_train,X_test,y_train,y_test = train_test_split(X,y,test_size = 0.3,random_state = 14)

#declaring parameters grid
k_value = list(range(2,11))
algorithm = ['auto','ball_tree','kd_tree','brute']
scores = []
best_comb = []
kfold = KFold(n_splits=5)

#hyperparameter tunning
for algo in algorithm:
  for k in k_value:
    knn = KNeighborsClassifier(n_neighbors=k,algorithm=algo)
    results = cross_val_score(knn,X_train,y_train,cv = kfold)

    print(f'Score:{round(results.mean(),4)} with algo = {algo} , K = {k}')
    scores.append(results.mean())
    best_comb.append((k,algo))

best_param = best_comb[scores.index(max(scores))]
print(f'
The Best Score : {max(scores)}')
print(f"['algorithm': {best_param[1]} ,'n_neighbors': {best_param[0]}]")

Score:0.6697 with algo = auto , K = 2
Score:0.6773 with algo = auto , K = 3
Score:0.7177 with algo = auto , K = 4
Score:0.734 with algo = auto , K = 5
Score:0.7017 with algo = auto , K = 6
Score:0.7417 with algo = auto , K = 7
Score:0.7017 with algo = auto , K = 8
Score:0.6533 with algo = auto , K = 9
Score:0.6613 with algo = auto , K = 10
Score:0.6697 with algo = ball_tree , K = 2
Score:0.6773 with algo = ball_tree , K = 3
Score:0.7177 with algo = ball_tree , K = 4
Score:0.734 with algo = ball_tree , K = 5
Score:0.7017 with algo = ball_tree , K = 6
Score:0.7417 with algo = ball_tree , K = 7
Score:0.7017 with algo = ball_tree , K = 8
Score:0.6533 with algo = ball_tree , K = 9
Score:0.6613 with algo = ball_tree , K = 10
Score:0.6697 with algo = kd_tree , K = 2
Score:0.6773 with algo = kd_tree , K = 3
Score:0.7177 with algo = kd_tree , K = 4
Score:0.734 with algo = kd_tree , K = 5
Score:0.7017 with algo = kd_tree , K = 6
Score:0.7417 with algo = kd_tree , K = 7
Score:0.7017 with algo = kd_tree , K = 8
Score:0.6533 with algo = kd_tree , K = 9
Score:0.6613 with algo = kd_tree , K = 10
Score:0.6697 with algo = brute , K = 2
Score:0.6773 with algo = brute , K = 3
Score:0.7177 with algo = brute , K = 4
Score:0.734 with algo = brute , K = 5
Score:0.7017 with algo = brute , K = 6
Score:0.7417 with algo = brute , K = 7
Score:0.7017 with algo = brute , K = 8
Score:0.6533 with algo = brute , K = 9
Score:0.6613 with algo = brute , K = 10

The Best Score : 0.7416666666666667
['algorithm': auto ,'n_neighbors': 7]

缺点：

1. 没办法确保得到最佳的参数组合。
2. 这是一个不断试错的过程，所以，非常的耗时。

2. 网格搜索

网格搜索是一种基本的超参数调优技术。它类似于手动调优，为网格中指定的所有给定超参数值的每个排列构建模型，评估并选择最佳模型。考虑上面的例子，其中两个超参数k_value =[2,3,4,5,6,7,8,9,10] & algorithm =[' auto '， ' ball_tree '， ' kd_tree '， ' brute ']，在这个例子中，它总共构建了9*4 = 36不同的模型。

from sklearn.model_selection import GridSearchCV

knn = KNeighborsClassifier()
grid_param = { 'n_neighbors' : list(range(2,11)) , 
              'algorithm' : ['auto','ball_tree','kd_tree','brute'] }
              
grid = GridSearchCV(knn,grid_param,cv = 5)
grid.fit(X_train,y_train)

#best parameter combination
grid.best_params_  #{'algorithm': 'auto', 'n_neighbors': 5}

#Score achieved with best parameter combination
grid.best_score_  #0.774

#all combinations of hyperparameters
grid.cv_results_['params']

#average scores of cross-validation
grid.cv_results_['mean_test_score']

至于为什么二者的结果会不一样，那是因为seed种子数也是一个超参数

缺点：

由于它尝试了超参数的每一个组合，并根据交叉验证得分选择了最佳组合，这使得GridsearchCV非常慢。

3. 随机搜索

使用随机搜索代替网格搜索的动机是，在许多情况下，所有的超参数可能不是同等重要的。随机搜索从超参数空间中随机选择参数组合，参数由n_iter给定的固定迭代次数的情况下选择。实验证明，随机搜索的结果优于网格搜索。

from sklearn.model_selection import RandomizedSearchCV

knn = KNeighborsClassifier()

grid_param = { 'n_neighbors' : list(range(2,11)) , 
              'algorithm' : ['auto','ball_tree','kd_tree','brute'] }

rand_ser = RandomizedSearchCV(knn,grid_param,n_iter=10)
rand_ser.fit(X_train,y_train)

#best parameter combination
rand_ser.best_params_  #{'n_neighbors': 7, 'algorithm': 'brute'}

#score achieved with best parameter combination
rand_ser.best_score_  #0.7256666666666667

#all combinations of hyperparameters
rand_ser.cv_results_['params']

#average scores of cross-validation
rand_ser.cv_results_['mean_test_score']

缺点：

随机搜索的问题是它不能保证给出最好的参数组合。

4. 贝叶斯搜索

贝叶斯优化属于一类优化算法，称为基于序列模型的优化(SMBO)算法。这些算法使用先前对损失f的观察结果，以确定下一个(最优)点来抽样f。该算法大致可以概括如下。

1. 使用先前评估的点X1*:n*，计算损失f的后验期望。
2. 在新的点X的抽样损失f，从而最大化f的期望的某些方法。该方法指定f域的哪些区域最适于抽样。

重复这些步骤，直到满足某些收敛准则。

from skopt import BayesSearchCV

import warnings
warnings.filterwarnings("ignore")

# parameter ranges are specified by one of below
from skopt.space import Real, Categorical, Integer

knn = KNeighborsClassifier()
#defining hyper-parameter grid
grid_param = { 'n_neighbors' : list(range(2,11)) , 
              'algorithm' : ['auto','ball_tree','kd_tree','brute'] }

#initializing Bayesian Search
Bayes = BayesSearchCV(knn , grid_param , n_iter=30 , random_state=14)
Bayes.fit(X_train,y_train)

#best parameter combination
Bayes.best_params_  #OrderedDict([('algorithm', 'ball_tree'), ('n_neighbors', 5)])

#score achieved with best parameter combination
Bayes.best_score_  #0.7741935483870968

#all combinations of hyperparameters
Bayes.cv_results_['params']

#average scores of cross-validation
Bayes.cv_results_['mean_test_score']

缺点：

要在2维或3维的搜索空间中得到一个好的代理曲面需要十几个样本，增加搜索空间的维数需要更多的样本。