机器学习模型训练测试完整步骤
文章的目的：这篇文章主要讲述模型的建立、以及测试的完整步骤，重点是各个步骤的关系以及作用。
文章的前提：这篇文章所有代码是用python编写，用knn算法对经典数据MNIST data（手写数字）进行建立模型。这是用一个实例说明，不用太在乎算法和数据本身。
文章结构：文章会解释每一步的作用，最后会贴出全部python代码，本文的数据下载地址：这里写链接内容
一、读取数据（原始数据）
读取数据可以看做数据收集，原始数据的状态。这里是csv文件，具体读取方式有很多，这里采用pandas的方法。
def opencsv(): # 使用pandas打开
data = pd.read_csv('data/train.csv')
data1 = pd.read_csv('data/test.csv')
train_x = data.values[0:, 1:] # 读入全部训练数据
train_y = data.values[0:, 0]
result_x = data1.values[0:, 0:] # 测试全部测试个数据
return train_x, train_y, result_x二、数据预处理
数据预处理处理是对数据提前进行处理和修正。主要包括：特征提取，特征降维、特征空值处理、特征转换（one-hot）、特征归一化；目标值空值处理，目标值转换（one-hot）
（其中常用的是：降维、空值、one-hot转换、归一化。）
（这里我的数据没有经过预处理，因为原始的数据是图片像素数据，具体每一种数据在sklearn上都有方法）
三、交叉验证数据划分
为了模型测试，先选择交叉验证方法，提前划分好数据。
（注意：交叉验证会改变数据顺序，若原始数据对你有用，可以先进行交叉验证，在进行降维等预处理，接下来会演示）
def data_pro(x,y):
x_train, x_test, y_train, y_test = train_test_split(x, y,test_size=0.1,random_state=33)
return x_train, x_test, y_train, y_test四、模型建立及测试
用处理好的数据建立训练模型，对模型的评价有很多参数，常用的有：得分（对的比例）、查准率、查全率、F1指数
#训练
knnClf = KNeighborsClassifier() # k=5 KNN中邻值为5，
knnClf.fit(x_train, ravel(y_train))
#预测
y_predict = knnClf.predict(x_test)
print("score on the testdata:",knnClf.score(x_test,y_test))
# print("score on the traindata:",knnClf.score(x_train,y_train))
print(classification_report(y_test,y_predict))五、预测的可能性计算
计算分类的概率大小
# 可能性
probablity = knnClf.predict_proba(x_test)
list_pro = []
for i in range(probablity.shape[0]):
pro = max(list(probablity[i]))
list_pro.append(pro)六、结果保存
将编号，原始结果，预测结果，预测概率保存csv
#输出
index = np.array(id).reshape((-1,1))[:,0:1]
result = pd.DataFrame(np.column_stack((index.reshape(-1,1),np.array(y_test).reshape(-1,1),np.array(y_predict).reshape(-1,1),np.array(list_pro).reshape(-1,1))),
columns=['ImageId','test_label','predict_lable','probablity'])
result.to_csv('result/knn_result.csv',index=False,header=True,encoding='gbk')七、错误分析
错误本身就是一个很重要的东西，将错误分类保存起来。当需要对具体错误类型分析的时候，可以逐个分析错误。（比如当对‘4’分类错误比价多的时候，可以加大‘4’的权重，使得其充分训练）
#错误分析
diff_index = []
for i in range(result.shape[0]):
diff_index.append(result['test_label'][i] != result['predict_lable'][i])
print(diff_index)
diff = result[diff_index]
diff_x = x_test_original[diff_index]
#查看每个错误
for i in range(len(diff_index)):
# print("label is:",diff['test_label'][i],"predict is:",diff['predict_lable'][i])
print("test label is :",diff.iloc[i]['test_label'],'predict label is :',diff.iloc[i]['predict_lable'])
x = diff_x[i]
img = x.reshape(28,28)
image_show(img)
diff.to_csv('result/knn_result_diff.csv',index=False,header=True,encoding='gbk')全部代码：
import pandas as pd
import time
from numpy import ravel, savetxt
from sklearn import svm
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import classification_report
from sklearn.model_selection import train_test_split
from sklearn.neighbors import KNeighborsClassifier
import numpy as np
from sklearn.decomposition import PCA
import matplotlib.pyplot as plt
def image_show(img):
plt.imshow(img)
plt.show()
def opencsv(): # 使用pandas打开
data = pd.read_csv('data/train.csv')
data1 = pd.read_csv('data/test.csv')
train_x = data.values[0:, 1:] # 读入全部训练数据
train_y = data.values[0:, 0]
result_x = data1.values[0:, 0:] # 测试全部测试个数据
return train_x, train_y, result_x
def data_pro(x,y):
x_train, x_test, y_train, y_test = train_test_split(x, y,test_size=0.1,random_state=33)
return x_train, x_test, y_train, y_test
def knnClassify(x_train, x_test, y_train, y_test):
id = range(1,x_test.shape[0]+1)
print("start run knn.")
#训练
knnClf = KNeighborsClassifier() # k=5 KNN中邻值为5，
knnClf.fit(x_train, ravel(y_train))
#预测
y_predict = knnClf.predict(x_test)
print("score on the testdata:",knnClf.score(x_test,y_test))
# print("score on the traindata:",knnClf.score(x_train,y_train))
print(classification_report(y_test,y_predict))
# 可能性
probablity = knnClf.predict_proba(x_test)
list_pro = []
for i in range(probablity.shape[0]):
pro = max(list(probablity[i]))
list_pro.append(pro)
#输出
index = np.array(id).reshape((-1,1))[:,0:1]
result = pd.DataFrame(np.column_stack((index.reshape(-1,1),np.array(y_test).reshape(-1,1),np.array(y_predict).reshape(-1,1),np.array(list_pro).reshape(-1,1))),
columns=['ImageId','test_label','predict_lable','probablity'])
result.to_csv('result/knn_result.csv',index=False,header=True,encoding='gbk')
#错误分析
diff_index = []
for i in range(result.shape[0]):
diff_index.append(result['test_label'][i] != result['predict_lable'][i])
print(diff_index)
diff = result[diff_index]
diff_x = x_test_original[diff_index]
#查看每个错误
for i in range(len(diff_index)):
# print("label is:",diff['test_label'][i],"predict is:",diff['predict_lable'][i])
print("test label is :",diff.iloc[i]['test_label'],'predict label is :',diff.iloc[i]['predict_lable'])
x = diff_x[i]
img = x.reshape(28,28)
image_show(img)
diff.to_csv('result/knn_result_diff.csv',index=False,header=True,encoding='gbk')
def svmClassify(train_x, train_y, test_x):
id = range(1, 28001)
t = time.time()
svc = svm.SVC(kernel='rbf', C=10)
svc.fit(train_x, train_y)
h = time.time()
print('time used:%f' % (h - t))
test_y = svc.predict(test_x)
k = time.time()
print('time used:%f' % (k - h))
savetxt('sklearn_svm_Result.csv', test_y, delimiter=',')
result = pd.DataFrame(np.column_stack((np.array(id).reshape((-1, 1))[:, 0:1], np.array(test_y).reshape((-1, 1))[:, 0:1])),
columns=['ImageId', 'Label'])
result.to_csv("sklearn_knn_Result2.csv", index=False, header=True, encoding='gbk')
if __name__ == "__main__":
print("start.")
#原数据
train_x_original, train_y_original, result_x_original = opencsv()
# 交叉验证
x_train_original, x_test_original, y_train, y_test = data_pro(train_x_original, train_y_original)
# 降维
pca = PCA(n_components=0.8, whiten=True)
train_x_pca = pca.fit_transform(x_train_original)
x_test_pca = pca.transform(x_test_original)
result_x_pca = pca.transform(result_x_original)
#knn
knnClassify(train_x_pca, x_test_pca, y_train, y_test)
#SVM
# svmClassify(train_x,train_y,test_x)
print("end.")