machine_learn/svm/svm_train.py

#!/usr/bin/python
# -*- encoding:utf-8 -*-
from sklearn import svm
import numpy as np
from sklearn.model_selection import GridSearchCV
import matplotlib.pyplot as plt


def read_data(path):
	with open(path) as f :
		lines=f.readlines()
	lines=[eval(line.strip()) for line in lines]
	X,y=zip(*lines)
	X=np.array(X)
	y=np.array(y)
	return X,y


def drawScatterAndLine(p, q, support):
    x1 = []
    x2 = []
    y1 = []
    y2 = []
    x3 = []
    y3 = []
    x4 = []
    y4 = []

    for idx,i in enumerate(q):
        item = (p[idx][0], p[idx][1])
        if item in support:
            is_supppot = True
            print(i, "support:", item)
        else:
            is_supppot = False
        if i == 0:
            if is_supppot:
                x3.append(p[idx][0])
                y3.append(p[idx][1])
            else:
                x1.append(p[idx][0])
                y1.append(p[idx][1])
        else:
            if is_supppot:
                x4.append(p[idx][0])
                y4.append(p[idx][1])
            else:
                x2.append(p[idx][0])
                y2.append(p[idx][1])


    plt.scatter(x1, y1)
    plt.scatter(x2, y2)
    plt.scatter(x3, y3, c="g")
    plt.scatter(x4, y4, c='black')

    plt.xlabel('p')
    plt.ylabel('q')
    plt.title('SVM')
    plt.show()


X_train,y_train=read_data("train_data")
X_test,y_test=read_data("test_data")

#  C对样本错误的容忍
#  默认核函数是rbf, 导致这个数据支撑向量不在图的边界点上
#  如果把核函数调成linear或者poly，支撑向量就在图的边界点上
model= svm.SVC(kernel='rbf')

model.fit(X_train, y_train)
print(model.support_vectors_)
print(model.support_)
print(len(model.support_))

drawScatterAndLine(X_train, y_train, model.support_vectors_)

score = model.score(X_test,y_test)
print(score)

#网格搜索
# search_space = {'C': np.logspace(-3, 3, 7)}
# print(search_space['C'])
# model= svm.SVC()
# gridsearch = GridSearchCV(model, param_grid=search_space)
# gridsearch.fit(X_train,y_train)
# cv_performance = gridsearch.best_score_
# test_performance = gridsearch.score(X_test, y_test)
# print("C",gridsearch.best_params_['C'])