import numpy as np,matplotlib.pyplot as plt,copy
import sklearn.datasets as dts
from sklearn.preprocessing import OneHotEncoder
#显示中文和负号
plt.rcParams['font.sans-serif']=['SimHei']
plt.rcParams['axes.unicode_minus']=False
def data_process():
# 9、加载wine数据集load_wine()
data=dts.load_wine()
# 10、获取特征x
x=data.data
# 11、获取标签y
y=data.target
# 12、标准化特征缩放
x_mean=np.mean(x,axis=0)
x_std=np.std(x,axis=0,ddof=1)
x=(x-x_mean)/x_std
# 13、初始化，特征添加全为1的第一列
m,n=x.shape
x=np.c_[np.ones((m,1)),x]
y=np.c_[y]
# 14、标签独热处理(新增)
one_hot=OneHotEncoder(categories='auto')#自动处理,不会从0-9编码
y=one_hot.fit_transform(y).toarray()#先训练再应用,toarray转化成矩阵
# 15、洗牌
np.random.seed(5)
order=np.random.permutation(m)
x=x[order]
y=y[order]
# 16、划分训练集：测试集=0.9：0.1
num=int(m*0.9)
train_x,test_x=np.split(x,[num])
train_y,test_y=np.split(y,[num])
return train_x,train_y,test_x,test_y
# 1、要求神经网络模型结构有四层（包括输入层，中间层2层，输出层）
# 2、要求中间层神经元个数满足9,8的结构，完成wine三分类神经网络底层实现
# 3、定义sigmoid函数及其导数
def sigmoid(g,deriy=False):
if deriy:
return g*(1-g)
else:
return 1.0/(1+np.exp(-g))
# 4、正向传播
def fp(x,theta1,theta2,theta3):
a1=x
z2=a1.dot(theta1)
a2=sigmoid(z2)
z3=a2.dot(theta2)
a3=sigmoid(z3)
z4=a3.dot(theta3)
a4=sigmoid(z4)
h=a4
return a2,a3,h
# 5、根据交叉熵，定义代价函数
def loss_func(h,y,lamda,theta1R,theta2,theta3):
m=len(h)
R=lamda/(2*m)*(np.sum(theta1R**2)+np.sum(theta2**2)+np.sum(theta3**2))
J=-np.mean(y*np.log(h)+(1-y)*np.log(1-h))+R
return J
# 6、进行反向传播return dt,s,
def bp(x,h,y,a2,a3,theta2,theta3,lamda,theta1R):
s4=h-y
s3=s4.dot(theta3.T)*sigmoid(a3,deriy=True)
s2=s3.dot(theta2.T)*sigmoid(a2,deriy=True)
dt3=a3.T.dot(s4)+lamda*theta3
dt2=a2.T.dot(s3)+lamda*theta2
dt1=x.T.dot(s2)+lamda*theta1R
return dt3,dt2,dt1
# 7、进行梯度下降
def train_model(x,y,alpha=0.7,iters=100,lamda=0.1):
xm,xn=x.shape
ym,yn=y.shape
loss_list=[]
theta1=np.random.randn(xn,9)
theta2=np.random.randn(9,8)
theta3=np.random.randn(8,yn)
for i in range(iters):
a2, a3, h=fp(x,theta1,theta2,theta3)
theta1R = copy.copy(theta1)
theta1R[0]=0
loss=loss_func(h,y,lamda,theta1R,theta2,theta3)
loss_list.append(loss)
dt3, dt2, dt1=bp(x,h,y,a2,a3,theta2,theta3,lamda,theta1R)
theta1=theta1-alpha*(1.0/xm)*dt1
theta2 = theta2 - alpha * (1.0 / xm) * dt2
theta3 = theta3 - alpha * (1.0 / xm) * dt3
return loss_list,theta1,theta2,theta3
# 8、定义精度函数，返回分类准确率
def acc_func(h,y):
y_label=np.argmax(y,axis=1)
h_label=np.argmax(h,axis=1)
acc=np.mean(y_label==h_label)
return acc
if __name__ == '__main__':
# 17、训练数据，调整超参数
train_x, train_y, test_x, test_y=data_process()
loss_list, theta1, theta2, theta3=train_model(train_x,train_y)
# 18、绘制代价函数图
plt.plot(loss_list,c='m')
plt.show()
# 19、输出训练集精度
a2, a3, train_h=fp(train_x,theta1,theta2,theta3)
print('训练集精度:',acc_func(train_h,train_y))
# 20、输出测试集精度
a2, a3, test_h = fp(test_x, theta1, theta2, theta3)
print('测试集精度:', acc_func(test_h, test_y))
# 23、取前两列特征，进行可视化展示
# 24、绘制样本真实分类情况
plt.scatter(test_x[:,1],test_x[:,2],c=np.argmax(test_y,axis=1),marker='o')
# 25、绘制预测分类情况，以星型*标注
plt.scatter(test_x[:,1],test_x[:,2],c=np.argmax(test_h,axis=1),marker='*')
plt.show()
