从零开始写代码-python解深度学习神经网络题目

题目来源：华为2020软挑热身赛

题目源文件：https://github.com/CatKungfu/huawei2020-LR

神经网络思路分析请看上一个解说视频：https://www.bilibili.com/video/BV1CK4y1A7jS/

本期博客视频版：https://www.bilibili.com/video/BV1xg411G7Dx/

 

 

 

 

 

 

 

#date:2021-05-31
#author:会武术的白猫
#theme:神经网络

import numpy as np

def read_train_data():#读取txt数据并整理成可以处理类型
    with open("data/train_data.txt",'r')as f:
        content=f.readlines()
    result_x=[]
    answer_Y=[]
    for item in content:
        jtem=item.split(',')
        answer_Y.append(int(jtem[-1]))
        jtem.pop(-1)
        result_x.append([float(i) for i in jtem])
    return [result_x,answer_Y]

def read_predict_data():#同上
    with open("data/test_data.txt",'r')as f:
        content=f.readlines()
    result_x=[]
    for item in content:
        jtem=item.split(',')
        result_x.append([float(i) for i in jtem])
    return result_x

def read_answer_data():#同上
    with open("data/answer.txt",'r')as f:
        content=f.readlines()
    answer_Y=[int(i) for i in content]
    return answer_Y

def sigmod(x):#激活函数计算
    return 1/(1+np.exp(-x))

def der_sigmod(x):#激活函数的导数计算
    return sigmod(x)*(1-sigmod(x))

def train():#训练函数，算出k权值矩阵和w权值矩阵
    res=read_train_data()
    result_x=res[0]
    answer_Y=res[1]
    n=len(result_x[0])
    error_min=0.001
    study=0.1
    nerve_num=8
    weight_k=np.random.random(size=(n,4))
    weight_w=np.random.random(size=(4,1))
    for i in range(len(result_x)):
        item=np.array(result_x[i])
        nerve_a=np.dot(item,weight_k).reshape(1,4)
        nerve_o=np.array([sigmod(j) for j in nerve_a]).reshape(1,4)
        y1=np.dot(nerve_o,weight_w)
        y2=sigmod(y1)
        L=(answer_Y[i]-y2)*(answer_Y[i]-y2)/float(2)
        if L<error_min:break
        mid=(y2-answer_Y[i])*der_sigmod(y1)
        weight_w-=study*mid*nerve_o.T
        mid_w=weight_w*(np.array([sigmod(j) for j in nerve_a]).T)
        weight_k-=study*mid*np.dot(mid_w,item.reshape(1,n)).T
        #return [weight_k,weight_w]
    return [weight_k,weight_w]

def predict():#读取数据给出预测二分类,计算正确率
    result_x=read_predict_data()
    res=train()
    weight_k=res[0]
    weight_w=res[1]
    nerve_a=np.dot(np.array(result_x),weight_k)
    nerve_o=np.array([sigmod(nerve_a[i][j]) for i in range(len(result_x)) for j in range(4)]).reshape(len(result_x),4)
    y1=np.dot(nerve_o,weight_w)
    y2=[sigmod(i) for i in y1]
    answer_Y=read_answer_data()
    num=0
    for i in range(len(answer_Y)):
        if (y2[i]>=0.5 and answer_Y[i]==1) or (y2[i]<0.5 and answer_Y[i]==0):
            num+=1
    result=num/float(len(result_x))
    print("正确率为{}%".format(result*100))

#read_train_data()
#read_predict_data()
#read_answer_data()
#train()
predict()