用sklearn进行手写数字的识别训练，获取学习率和惩罚系数，机器学习入门推荐案例（1）

理论总是让人感觉乏味和枯燥，对于学习率和惩罚系数是如何确定的呢？我们可以用列举法来评价每个参数的得分，进而获取最优的模型参数。

一、 获取手写数据集合

from sklearn.datesets import load_digits
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
import matplotlib.pyplot as plt
def load_data():
	data = load_digits()
	x, y = data.data, data.target
	#切割测试集和训练集
	train_x, test_x, train_y, test_y = train_test_split(test_size= 0.3, random_state=10)
	#标准化数据
	ss = StandardScaler()
	ss.fit(train_x)
	train_x = ss.transform(train_x)
	test_x = ss.transform(test_x)	
	return train_x, test_x, train_y, test_y;

处理完数据后，可以展示一下数据的样式，让我们看看图片张啥样

def show_img(X):
	images = X.reshape(-1, 8, 8 ) # 格式为（-1， 64,）图片格式为8 * 8
	# 构建绘图
	fig, axi = plt.subslots(3, 5)
	for i, ax in enumerate(axi.flat):
		img = images[i]
		ax.imshow(img, cmap='gray')
		ax.set(xticks=[], yticks=[])
	plt.tight_layout()
	plt.show()

开始寻找最优的学习率和惩罚系数

def model_fit(X,y, k=5):
	learning_rates = [0.001, 0.003, 0.01, 0.03, 0.1, 0.3, 1, 3]
    penalties = [0, 0.03, 0.01, 0.1, 0.3, 1, 3] #惩罚系数
    all_models = []
    for learning_rate in learning_rates:
    	for penalty in penalties:
    	print('正训练模型:下降参数为%s，正则化为%s'%(learning_rate, penalty))
    	#建立逻辑回归模型
        model_score = []
	    #运用交叉验证法取样
	    kf = KFold(n_split= k, shuffle= True, random_state = 20)
	    model = SGDclassifier(loss = 'log', learning_rate = "constant", penalty = "l2", eta0 = learning_rate, alpha = penalty)
	    for train_idx, dev_idx in kf.split(X):
	    	x_train, x_dev = X[train_idx], x[dev_idx]
	    	y_train, y_dev = y[train_idx], y[dev_idx]
	    	model.fit(x_train, x_dev)
	    	s = model.score(y_train, y_dev)
	    	model_score.append(s)
	    all_models.append([np.mean(model_score), learning_rate, penalty])
	 print("模型最优解：", sorted(all_models, reverse= True)[0]
	 

测试运行

if __name__ == "__main__":
    train_x, test_x, train_y, test_y = load_data()
        
    print(train_x.shape)
    print(train_y.shape)
    
    # model_fit(train_x, train_y)
    
    model = SGDClassifier(loss='log', penalty='l2',learning_rate="constant", alpha=0.01, eta0=0.01)
    model.fit(train_x, train_y)
    y_pred = model.predict(test_x)
    print(test_y)
    print(y_pred)

完整源码：

# -*- coding: utf-8 -*-

"模型的正则化和超参数的获取"

#1 引入需要的包 
import numpy as np 
from sklearn.datasets import load_digits
from sklearn.model_selection import train_test_split, KFold
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import SGDClassifier
#2 划分数据为训练和测试

def load_data():
    data = load_digits()
    x, y = data.data, data.target
    train_x, test_x, train_y, test_y = train_test_split(x, y , test_size = 0.3, random_state = 20)
#3 标准化数据
    ss = StandardScaler()
    ss.fit(train_x)
    train_x = ss.transform(train_x)
    test_x = ss.transform(test_x)
    return train_x, test_x, train_y, test_y
#4 进行模型的验证取最佳的超参数和正则化参数

def model_fit(x, y, k = 5):
    learning_rates = [0.001, 0.003, 0.01, 0.03, 0.1, 0.3, 1, 3]
    penalties = [0, 0.03, 0.01, 0.1, 0.3, 1, 3]
    all_models = []
    for learning_rate in learning_rates:
        for penalty in penalties:
            print('正训练模型:下降参数为%s，正则化为%s'%(learning_rate, penalty))
            #建立逻辑回归模型
            model_score = []
            model = SGDClassifier(loss='log', penalty='l2', learning_rate='constant', eta0=learning_rate, alpha=penalty)
            kf = KFold(n_splits=k, shuffle=True, random_state=10)
            for train_idx, dev_idx in kf.split(x):
                x_train, x_dev = x[train_idx], x[dev_idx]
                y_train, y_dev = y[train_idx], y[dev_idx]
                model.fit(x_train, y_train)
                s = model.score(x_dev, y_dev)
                model_score.append(s)
            all_models.append([np.mean(model_score), learning_rate, penalty])
    print('最优模型：', sorted(all_models, reverse= True)[0])

if __name__ == "__main__":
    train_x, test_x, train_y, test_y = load_data()
        
    print(train_x.shape)
    print(train_y.shape)
    
    # model_fit(train_x, train_y)
    
    model = SGDClassifier(loss='log', penalty='l2',learning_rate="constant", alpha=0.01, eta0=0.01)
    model.fit(train_x, train_y)
    y_pred = model.predict(test_x)
    print(test_y)
    print(y_pred)