深度学习-神经网络实战

红酒的质量预测：https://www.kaggle.com/uciml/red-wine-quality-cortez-et-al-2009/tasks?taskId=4684

ActivateFuncs.py

import pandas as pd
import os
import numpy as np
def ReLu(x):
    return np.maximum(0,x)

def deriv_ReLu(x):
    ret = x
    ret[ret<=0]=0
    ret[ret>0]=1
    return ret

main.py

import pandas as pd
import os
import numpy as np
import matplotlib as mpl
import  matplotlib.pyplot as  plt
import ActivateFuncs
from sklearn.preprocessing import StandardScaler

#参数设置
hidden_layer_number = 5 #隐藏层数
hidden_neuron_number = 11 #隐藏层神经元数
feature_number = 11 #特征数
learn_rate = 0.01 #学习率
max_epochs = 100 #循环迭代次数
sample_number = 1600#样本数
mse_final = 6.5e-4  #设置一个均方误差的阈值，小于它则停止迭代

#读取数据
trainFile = "C:\\Users\Mr.K\\Desktop\\机器学习：红酒质量预测\\winequality-red.csv"
data = pd.read_csv(trainFile);#读取到的是dataframe类型
data = np.array(data)
traindata = data[slice(0,1600)]

#矩阵化
inputs = traindata[0:sample_number,0:11].T  #输入数据
y_hat = traindata[0:sample_number,11:12].T  #真实质量值

ss=StandardScaler()
inputs = ss.fit_transform(inputs)

#从0开始
W = []
B = []
#初始化W,B
for i in range(0,hidden_layer_number):
    W.append(np.random.randn(11, 11) * 0.05)
    B.append(np.ones((11,1)))
#输出层W
W.append(np.random.randn(1,11)*0.05)
B.append(np.ones((1,1)))


#从1开始，Z[1]表示第一层隐藏层
Z = [np.zeros((hidden_neuron_number,1))]
A = [inputs]
for layer in range(0,hidden_layer_number):#初始化
    Z.append(np.zeros((hidden_neuron_number,1)))
    A.append(np.zeros((hidden_neuron_number,1)))
Z.append(np.array([[0]]))
A.append(np.array([[0]]))

mse_history=[]  #空列表，存储迭代的误差
dw = []
db = []
dz = []
err0 = 0
for i in range(0,max_epochs):
    for layer in range(1,hidden_layer_number+1):
        Z[layer] = np.dot(W[layer-1],A[layer-1])+B[layer-1]
        A[layer] = ActivateFuncs.ReLu(Z[layer])
    #输出层传播
    Z[hidden_layer_number+1] = np.dot(W[hidden_layer_number], A[hidden_layer_number]) + B[hidden_layer_number]
    A[hidden_layer_number+1] = ActivateFuncs.ReLu(Z[hidden_layer_number+1])

    err = np.average(np.square(y_hat - A[hidden_layer_number + 1]))
    mse_history.append(err)

    #反向传播
    dz = A[hidden_layer_number+1]-y_hat

    for layer in range(hidden_layer_number+1,0,-1):
        dw = np.dot(dz , A[layer-1].T) / sample_number
        db = np.sum(dz,axis=1,keepdims=True) / sample_number
        dz = np.dot(W[layer-1].T , dz) * ActivateFuncs.deriv_ReLu(Z[layer-1])

        W[layer-1] -= learn_rate * dw
        B[layer-1] -= learn_rate * db

    err = np.average(np.square(y_hat-A[hidden_layer_number+1]))
    if i==0: err0 = err
    print('第%d次训练完成！ 错误率：'%(i)+str(err))

print('错误率下降'+str(err0-err))

#损失值画图
print(mse_history)
loss=np.log10(mse_history)
min_mse=min(loss)
plt.plot(loss,label='loss')
plt.plot([0,len(loss)],[min_mse,min_mse],label='min_mse')
plt.xlabel('iteration')
plt.ylabel('MSE loss')
plt.title('Log10 MSE History',fontdict={'fontsize':18,'color':'red'})
plt.legend()
plt.show()