Elman神经网络

import numpy as np
from matplotlib import pyplot as plt
########################## 设置模型参数 ############################

""" 
    训练轮次： 500
    输入层节点数： 3
    隐藏层节点数： 4
    输出层节点数： 1 
"""

epochs = 500
hidden_size, output_size = 4, 1
lr = 0.005

############################ 数据准备 ##############################

""" 
    输入数据: X
    目标函数: y 
"""

X = np.array([[1, 1, -1, 1],
              [1, -1, -1, 1],
              [-1, 1, -1, 1],
              [-1, -1, -1, 1]])

target = np.array([1, 0, 0, 1]).reshape(4, 1)

# 初始化承接层
context = np.zeros((X.shape[0], hidden_size))
input = np.hstack((context, X))

# 初始化权重
w_input = np.random.rand(X.shape[0], hidden_size + X.shape[1])

w_out = np.random.rand(hidden_size, 1)

############################ 搭建模型 #############################

# 激励函数


def sigmoid(x):
    return 1/(1+np.exp(-x))

# 正向传播过程


def forward(x, w1, w2):
    xw = np.dot(x, w1.T)
    context = sigmoid(xw)

    out = np.dot(context, w2)
    return context, out

# 反向传播过程

def backward(dout, contxt, w1, w2, x):
    d_contxt = np.dot(dout, w2.T)
    dw2 = np.dot(contxt, dout)         

    dxw = d_contxt * (1.0 - context) * context
    dw1 = np.dot(dxw, x)
    return dw1, dw2


for i in range(epochs):
    # 预测
    context, out = forward(input, w_input, w_out)

    # 计算损失
    loss = np.square(target - out) / X.shape[0]

    # 反向传播
    dw1, dw2 = backward(loss, context, w_input, w_out, input)

    # 更新
    w_input -= lr * dw1
    w_out -= lr * dw2

    # 评价指标
    print('-----------------------Epoch is :[{}/{}]-----------------------'.format(i+1, epochs))
    print("out: {} 
Loss: {}".format(out.T, loss.T))

　　

该程序只是简单的写了一个框架，希望能大家提供一种编程思维。