简单的GRU实例代码

import numpy as np
# 定义sigmoid函数
def sigmoid(x):
    return 1 / (1 + np.exp(-x))
class RNN:
    def __init__(self, input_size, hidden_size, output_size):
        # 设定超参数
        self.input_size = input_size
        self.hidden_size = hidden_size
        self.output_size = output_size

        # 初始化权重和偏置
        self.Wxh = np.random.randn(hidden_size, input_size) * 0.01  # 输入到隐藏层的权重
        self.Whh = np.random.randn(hidden_size, hidden_size) * 0.01  # 隐藏层到隐藏层的权重
        self.Why = np.random.randn(output_size, hidden_size) * 0.01  # 隐藏层到输出层的权重
        self.bh = np.zeros((hidden_size, 1))  # 隐藏层的偏置
        self.by = np.zeros((output_size, 1))  # 输出层的偏置

    def forward(self, inputs):
        # 初始化隐藏状态和输出
        self.h = np.zeros((self.hidden_size, 1))
        self.outputs = []
        for x in inputs:
            # 更新隐藏状态
            self.h = np.tanh(np.dot(self.Wxh, x) + np.dot(self.Whh, self.h) + self.bh)
            # 计算输出
            y = np.dot(self.Why, self.h) + self.by
            # 应用sigmoid激活函数
            output = sigmoid(y)
            self.outputs.append(output)
        return self.outputs

    def backward(self, inputs, targets, learning_rate=0.1):
        # 初始化梯度
        dWxh = np.zeros_like(self.Wxh)
        dWhh = np.zeros_like(self.Whh)
        dWhy = np.zeros_like(self.Why)
        dbh = np.zeros_like(self.bh)
        dby = np.zeros_like(self.by)
        dh_next = np.zeros_like(self.h)

        for i in reversed(range(len(inputs))):
            # 计算输出误差
            dy = self.outputs[i] - targets[i]
            # 计算输出层的梯度
            dWhy += np.dot(dy, self.h.T)
            dby += dy
            # 计算隐藏层的误差
            dh = np.dot(self.Why.T, dy) + dh_next
            # 应用tanh的导数
            dh_raw = (1 - self.h ** 2) * dh
            # 计算隐藏层的梯度
            dWxh += np.dot(dh_raw, inputs[i].T)
            dWhh += np.dot(dh_raw, self.h.T)
            dbh += dh_raw
            # 更新dh_next
            dh_next = np.dot(self.Whh.T, dh_raw)

        for dparam in [dWxh, dWhh, dWhy, dbh, dby]:
            np.clip(dparam, -5, 5, out=dparam)  # 防止梯度爆炸

        # 更新权重和偏置
        self.Wxh -= learning_rate * dWxh
        self.Whh -= learning_rate * dWhh
        self.Why -= learning_rate * dWhy
        self.bh -= learning_rate * dbh
        self.by -= learning_rate * dby


# 测试代码
# 定义数据和标签
inputs = [np.array([[1], [0], [1]]), np.array([[0], [1], [0]])]
targets = [np.array([[1]]), np.array([[0]])]

input_size = inputs[0].shape[0]
hidden_size = 25
output_size = targets[0].shape[0]

# 创建RNN模型,并进行训练
rnn = RNN(input_size, hidden_size, output_size)
for epoch in range(1000):
    outputs = rnn.forward(inputs)
    loss = np.mean((np.array(outputs)-np.array(targets)) ** 2)
    rnn.backward(inputs, targets)
    if (epoch + 1) % 100 == 0:
        print("次数:", epoch + 1, "误差:", loss)

# 在新数据上进行预测
new_input = np.array([[1], [1], [1]])
output = rnn.forward([new_input])
print("输入:", new_input.flatten())
print("输出:", output)

评论

h
hw_8088

这个代码可以运行,就是不知道逻辑有没有错误

赞 0