BP神经网络算法 Python实现

Norman Mises

已于 2022-10-25 22:49:30 修改

阅读量1.8k

点赞数 3

文章标签： python 神经网络算法

于 2022-10-25 22:43:18 首次发布

本文链接：https://blog.csdn.net/weixin_43212810/article/details/127523466

版权

BP神经网络算法
实验要求：

掌握BP神经网络算法的实现
掌握利用BP神经网络对样本数据分类，并观察训练时间及迭代次数对结果的影响

Python 代码

# -*- coding: utf-8 -*-
# @Time    : 2022/10/25 22:48
# @Author  : Norman Mises
# @FileName: train.py
# @Software: PyCharm

import math
import time

import numpy as np
import pandas as pd
import yaml


def train(dat, id):
    x = dat
    global d
    d = [-1 for _ in range(num_layers[-1])]  # 希望输出值
    d[id] = 1
    # print(d)

    global y
    y[0] = x
    for i in range(1, n):  # 每一层
        for j in range(num_layers[i]):  # 每一个
            net = sum([weights[i][j][k] * y[i - 1][k] for k in range(num_layers[i - 1])])
            f = lambda net_: 1 / (1 + math.e ** ((theta[i][j] - net_) / h0))
            y[i][j] = f(net)

    # for i in y[-1]:
    #     print(f'{i:.6f}', end=',')

    Ep = 0.5 * sum([(d[i] - y[-1][i]) ** 2 for i in range(num_layers[-1])])
    # print(f'{Ep}')
    if Ep > 0.50001:
        return False
    else:
        return True


def correct():
    global delta
    for i in range(n - 1, 0, -1):  # 每一层
        for j in range(num_layers[i]):
            if i == n - 1:
                delta[i][j] = y[i][j] * (1 - y[i][j]) * (d[j] - y[-1][j])
            else:
                delta[i][j] = y[i][j] * (1 - y[i][j]) * sum(
                    [delta[i + 1][q] * weights[i + 1][q][j] for q in range(num_layers[i + 1])])
            for k in range(num_layers[i - 1]):
                weights[i][j][k] += eta * delta[i][j] * y[i - 1][k]


def test(x):
    out = y
    out[0] = x
    for i in range(1, n):  # 每一层
        for j in range(num_layers[i]):  # 每一个
            net = sum([weights[i][j][k] * out[i - 1][k] for k in range(num_layers[i - 1])])
            f = lambda net_: 1 / (1 + math.e ** ((theta[i][j] - net_) / h0))
            out[i][j] = f(net)

    for i in range(num_layers[-1]):
        print(f'第 {i} 类：{out[-1][i]:.6f}')


if __name__ == "__main__":
    with open('testcases/training_data.yaml', 'r', encoding='utf-8') as f:
        data_yaml = yaml.load(f, yaml.FullLoader)
        print(data_yaml['data'][0])

    label = data_yaml['label'][0]
    df = pd.DataFrame(data_yaml['data'][0])
    data = (df - df.min()) / (df.max() - df.min())
    data = data.values
    print(data)
    print(label)

    h0 = float(input(f'h0: '))
    eta = float(input(f'eta: '))
    n = int(input('层数：'))
    num_layers = []
    weights = []
    theta = []  # 神经元阈值
    delta = []
    d = []
    y = []
    print('每层神经元数：')
    for i in range(n):
        num_layers.append(int(input(f'第 {i} 层：')))
        theta.append(np.random.rand(num_layers[i]))
        delta.append(np.zeros(num_layers[i]))
        y.append(np.zeros(num_layers[i]))
        # y.append([] for _ in range(num_layers[i]))
        if i > 0:
            weights.append(np.random.rand(num_layers[i], num_layers[i - 1]))
        else:
            weights.append([])

    print(f'每层神经元与其低一层的连接权：')
    for i in weights:
        print(i)

    print(f'神经元阈值：')
    for i in theta:
        print(i)
    times = int(input(f'训练次数上限（-1 为无上限）：'))
    start = time.time()
    cnt = 1
    while not train(data[cnt % len(data)], label[cnt % len(data)]):
        if times != -1 and cnt >= times:
            break
        cnt += 1
        correct()
    end = time.time()

    print(f'训练结束后每层神经元与其低一层的连接权：')
    for i in weights:
        print(i)
    print(f'训练次数：{cnt}')
    print(f'训练时间：{(end - start):.6f} s')

    try:
        while True:
            print(f'输入测试数据：')
            t = input().split(',')
            t = [float(i.strip()) for i in t]
            print(t)
            test(t)
    except:
        print('测试结束')