基于Python环神经网络预测正弦信号与短句源码+项目运行说明+详细注释.zip

import numpy as np from module import Layers class RNN(Layers): """ z_t=x_t @ U + h_(t-1) @ W + B h_t = tanh(z_t) 输入为x=[x_1, x_2, x_3 . .],形状为input_length*input_size 输出为h=[h_1, h_2, h_3 . .],形状为input_length*hidden_size """ def __init__(self, name, input_size, hidden_size): """ 循环神经网络初始化 Parameters ---------- name:str 循环神经网络名称 input_size:int 输入向量维度，RNN输入数据是形状为input_length*input_size的二维数组 hidden_size:int 隐藏层向量维度，RNN输出数据是形状为input_length*hidden_size的二维数组 """ super(RNN, self).__init__(name) self.input_size = input_size self.hidden_size = hidden_size self.U = np.random.randn(self.input_size, self.hidden_size) self.W = np.random.randn(self.hidden_size, self.hidden_size) self.B = np.random.randn(self.hidden_size) self.h_list = np.array([0]) self.x_list = np.array([0]) self.grad_U = np.zeros((self.input_size, self.hidden_size)) self.grad_W = np.zeros((self.hidden_size, self.hidden_size)) self.grad_B = np.zeros((self.hidden_size,)) def forward(self, x, h): """ RNN神经网络前向计算公式 Parameters ---------- x:numpy.ndarray 输入数据，二维矩阵，大小为input_length*input_size h:numpy.ndarray 输入隐藏层数据，二维矩阵，大小为1*hidden_size Returns ------- h_list:numpy.ndarray 输出数据，二维矩阵，大小为input_length*output_size h:numpy.ndarray 输出隐藏层数据，二维矩阵，大小为1*hidden_size """ self.x_list = x self.h_list = np.zeros((self.x_list.shape[0], self.hidden_size)) for i in range(self.x_list.shape[0]): z = self.x_list[i] @ self.U + h @ self.W + self.B h = np.tanh(z) self.h_list[i] = h return self.h_list, h def backward(self, grad_in): """ RNN神经网络误差反向传播计算 Parameters ---------- grad_in:numpy.ndarray 输入误差梯度，二维矩阵，大小为input_length*hidden_size Returns ------- grad_out:numpy.ndarray 输出误差梯度，二维矩阵，大小为input_length*input_size """ grad_out = np.zeros_like(self.x_list) h_buf = (1 - self.h_list ** 2) grad_z = grad_in[-1:] * h_buf[-1:] grad_out[-1:] = grad_z @ self.U.T self.grad_U = self.x_list[-1:].T @ grad_z self.grad_W = self.h_list[-2:-1].T @ grad_z self.grad_B = grad_z[0] for t in range(self.h_list.shape[0] - 2, 0, -1): grad_z = (grad_in[t:t + 1] + (grad_z @ self.W.T)) * h_buf[t:t + 1] grad_out[t:t + 1] = grad_z @ self.U.T self.grad_U += self.x_list[t:t + 1].T @ grad_z self.grad_W += self.h_list[t - 1:t].T @ grad_z self.grad_B += grad_z[0] grad_z = (grad_in[0:1] + (grad_z @ self.W.T)) * h_buf[0:1] grad_out[0:1] = grad_z @ self.U.T self.grad_U += self.x_list[0:1].T @ grad_z self.grad_W += 0 self.grad_B += grad_z[0] for i in [self.grad_U, self.grad_W, self.grad_B]: np.clip(i, -5, 5, out=i) return grad_out def update(self, lr=1e-3): """ 参数更新 Parameters ---------- lr:float 参数更新步长 """ self.U -= lr * self.grad_U self.W -= lr * self.grad_W self.B -= lr * self.grad_B def get_param(self): param = {self.name + '-U': self.U, self.name + '-W': self.W, self.name + '-B': self.B} return param def set_param(self, param): # 获取神经网络所有参数，避免传入字典的键值缺失 d = self.get_param() # 更新参数字典 d.update(param) self.U = d[self.name + '-U'] self.W = d[self.name + '-W'] self.B = d[self.name + '-B']