ML工作流中最困難的部分之一是為模型找到最好的超參數。ML模型的性能與超參數直接相關。超參數調優的越好，得到的模型就越好。調優超參數可能是非常乏味和困難的，更像是一門藝術而不是科學。
超參數
超參數是在建立模型時用於控制算法行為的參數。這些參數不能從常規訓練過程中獲得。在對模型進行訓練之前，需要對它們進行賦值。
1. 傳統手工搜索
在傳統的調參過程中，我們通過訓練算法手動檢查隨機超參數集，並選擇符合我們目標的最佳參數集。
我們看看代碼：
#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'\nThe 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]
缺點：
- 沒辦法確保得到最佳的參數組合。
- 這是一個不斷試錯的過程，所以，非常的耗時。
2. 網格搜索
網格搜索是一種基本的超參數調優技術。它類似於手動調優，為網格中指定的所有給定超參數值的每個排列構建模型，評估並選擇最佳模型。考慮上面的例子，其中兩個超參數k_value =[2,3,4,5,6,7,8,9,10] & algorithm =[' auto '， ' ball_tree '， ' kd_tree '， ' brute ']，在這個例子中，它總共構建了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']
至於為什么二者的結果會不一樣，那是因為seed種子數也是一個超參數
缺點：
由於它嘗試了超參數的每一個組合，並根據交叉驗證得分選擇了最佳組合，這使得GridsearchCV非常慢。
3. 隨機搜索
使用隨機搜索代替網格搜索的動機是，在許多情況下，所有的超參數可能不是同等重要的。隨機搜索從超參數空間中隨機選擇參數組合，參數由n_iter給定的固定迭代次數的情況下選擇。實驗證明，隨機搜索的結果優於網格搜索。
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']
缺點：
隨機搜索的問題是它不能保證給出最好的參數組合。
4. 貝葉斯搜索
貝葉斯優化屬於一類優化算法，稱為基於序列模型的優化(SMBO)算法。這些算法使用先前對損失f的觀察結果，以確定下一個(最優)點來抽樣f。該算法大致可以概括如下。
- 使用先前評估的點X1*:n*，計算損失f的后驗期望。
- 在新的點X的抽樣損失f，從而最大化f的期望的某些方法。該方法指定f域的哪些區域最適於抽樣。
重復這些步驟，直到滿足某些收斂准則。
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']
缺點：
要在2維或3維的搜索空間中得到一個好的代理曲面需要十幾個樣本，增加搜索空間的維數需要更多的樣本。