支持向量回歸機（SVR）代碼

SVR的代碼（python）

項目中一個早期版本的代碼，PCA-SVR，參數尋優采用傳統的GridsearchCV。

from sklearn.decomposition import PCA
from sklearn.svm import SVR
from sklearn.model_selection import train_test_split
from sklearn.model_selection import GridSearchCV
from sklearn.metrics import r2_score, mean_squared_error, mean_absolute_error
from sklearn.preprocessing import StandardScaler, MinMaxScaler
from numpy import *
import numpy as np
import matplotlib.pyplot as plt
import xlrd
from svmutil import *
import pandas as pd

'''前言'''
# pca - svr
# CG測試

'''預設參數'''
fname = "all01.xlsx"  # 訓練數據文件讀取 26hao
random_1 = 34  # 樣本集選取隨機種子
random_2 = 4  # 訓練集選取隨機種子
newpca = 6  # 降維
yuzhi = 50  # 異常點閾值
rate_1 = 0.8  # 樣本集驗證集
rate_2 = 0.8  # 訓練集測試集
bestc = 384  # c
bestg = 9  # gamma

'''數據讀取'''
# xlrd生成對excel表進行操作的對象
...

# 輸入輸出分割
data_x = data[:, 1:11]
data_y = data[:, 0:1]

'''PCA'''
pca = PCA(n_components=newpca)  # 加載PCA算法，設置降維后主成分數目為
data_x = pca.fit_transform(data_x)  # 對樣本進行降維
print(pca.components_)  # 輸出主成分，即行數為降維后的維數，列數為原始特征向量轉換為新特征的系數
print(pca.explained_variance_ratio_)  # 新特征 每維所能解釋的方差大小在全方差中所占比例

'''數據划分'''
# 樣本數據分割
train_data_x, predict_data_x, train_data_y, predict_data_y = train_test_split(data_x, data_y, test_size=rate_1,
                                                                              random_state=random_1)

# 訓練數據分割
train_x, test_x, train_y, test_y = train_test_split(train_data_x, train_data_y, test_size=rate_2, random_state=random_2)
predict_x = predict_data_x
predict_y = predict_data_y

# reshape y
test_y = np.reshape(test_y, -1)
train_y = np.reshape(train_y, -1)
predict_y = np.reshape(predict_y, (-1, 1))

# StandardScaler x
ss_X = StandardScaler()
ss_X.fit(train_data_x)  # 20%
train_x = ss_X.transform(train_x)
test_x = ss_X.transform(test_x)
predict_x = ss_X.transform(predict_x)

'''參數優化與SVR'''
# 網格搜索交叉驗證（GridSearchCV）：以窮舉的方式遍歷所有可能的參數組合
# 測試用
# param_grid = {'gamma': [bestg], 'C': [bestc]}
# rbf_svr_cg = GridSearchCV(SVR(kernel='rbf'), param_grid, cv=5)
# rbf_svr_cg.fit(train_x,train_y)
# bestc = rbf_svr_cg.best_params_.get('C')
# bestg = rbf_svr_cg.best_params_.get('gamma')

# 最優參數
print(bestc, bestg)
param_grid = {'gamma': [bestg], 'C': [bestc]}
rbf_svr = SVR(kernel='rbf',param_grid) # 需要修改

# 訓練
rbf_svr.fit(train_x, train_y)

# 預測
test_y_predict = rbf_svr.predict(test_x)
test_y_predict = np.reshape(test_y_predict, (-1, 1))
predict_y_predict = rbf_svr.predict(predict_x)
predict_y_predict = np.reshape(predict_y_predict, (-1, 1))

'''去異常點'''
print('樣本集：', len(train_data_y))
print('驗證集：', len(predict_data_y))
size = len(test_y_predict)
count = 0
for i in range(size):
    if abs(test_y_predict[size - i - 1] - test_y[size - i - 1]) > yuzhi:
        test_y_predict = np.delete(test_y_predict, size - i - 1)
        test_y = np.delete(test_y, size - i - 1)
        count = count + 1
print('測試集異常點', count)
size = len(predict_y_predict)
count = 0
for i in range(size):
    if abs(predict_y_predict[size - i - 1] - predict_y[size - i - 1]) > yuzhi:
        predict_y_predict = np.delete(predict_y_predict, size - i - 1)
        predict_y = np.delete(predict_y, size - i - 1)
        count = count + 1
print('驗證集異常點', count)

'''評估'''
# # 使用r2__score模塊，並輸出評估結果，擬合程度，R2決定系數，衡量模型預測能力好壞（真實與預測的相關程度百分比）
# print('The value of R-squared of kernal=rbf is',r2_score(test_y,test_y_predict))
# # 使用mean_squared_error模塊，輸出評估結果，均方誤差
# print('The mean squared error of kernal=rbf is',mean_squared_error(test_y,test_y_predict))
# # 使用mean_absolute_error模塊，輸出評估結果，平均絕對誤差
# print('The mean absolute error of kernal=rbf is',mean_absolute_error(test_y,test_y_predict))

# 使用r2__score模塊，並輸出評估結果，擬合程度，R2決定系數，衡量模型預測能力好壞（真實與預測的相關程度百分比）
print('The value of R-squared of kernal=rbf is', r2_score(predict_y, predict_y_predict))
# 使用mean_squared_error模塊，輸出評估結果，均方誤差
print('The mean squared error of kernal=rbf is', mean_squared_error(predict_y, predict_y_predict))
# 使用mean_absolute_error模塊，輸出評估結果，平均絕對誤差
print('The mean absolute error of kernal=rbf is', mean_absolute_error(predict_y, predict_y_predict))
# r
X1 = pd.Series(np.reshape(predict_y,-1))
Y1 = pd.Series(np.reshape(predict_y_predict,-1))
print('The r is', X1.corr(Y1, method="pearson"))
print('The r is', sqrt(r2_score(predict_y, predict_y_predict)))

'''作圖'''
# PRN
print('PRN:', fname)

# PCA
print()

# 殘差
diff_predict = predict_y_predict - predict_y
plt.plot(diff_predict, color='black', label='error')
plt.xlabel("no.")
plt.ylabel("error(m)")
plt.title('xxx')
plt.grid()
plt.legend()
plt.show()

# 真實/模型_1
plt.plot(predict_y, color='g', label='dtu15mss')
plt.plot(predict_y_predict, color='b', label='pre')
plt.xlabel("xxx")
plt.ylabel("error(m)")
plt.title('xxx')
plt.grid()
plt.legend()
plt.show()

# 真實/模型_2
fig = plt.figure(3)
ax1 = fig.add_subplot(2, 1, 1)
ax1.plot(predict_y, color='g', label='dtu15mss')
ax2 = fig.add_subplot(2, 1, 2)
ax2.plot(predict_y_predict, color='b', label='pre')
plt.show()

# 真實/模型_3
p_x = [x for x in range(int(min(predict_y)) - 5, int(max(predict_y)) + 5)]
p_y = p_x
plt.plot(p_x, p_y, color='black', label='1')
plt.scatter(predict_y_predict, predict_y, s=10, color='b', marker='x',
            label='0')  # https://www.cnblogs.com/shanlizi/p/6850318.html
plt.xlabel('PRE')
plt.ylabel('DTU')
plt.show()