简单的GRU实例代码

1import numpy as np
2# 定义sigmoid函数
3def sigmoid(x):
4    return 1 / (1 + np.exp(-x))
5class RNN:
6    def __init__(self, input_size, hidden_size, output_size):
7        # 设定超参数
8        self.input_size = input_size
9        self.hidden_size = hidden_size
10        self.output_size = output_size
11
12        # 初始化权重和偏置
13        self.Wxh = np.random.randn(hidden_size, input_size) * 0.01  # 输入到隐藏层的权重
14        self.Whh = np.random.randn(hidden_size, hidden_size) * 0.01  # 隐藏层到隐藏层的权重
15        self.Why = np.random.randn(output_size, hidden_size) * 0.01  # 隐藏层到输出层的权重
16        self.bh = np.zeros((hidden_size, 1))  # 隐藏层的偏置
17        self.by = np.zeros((output_size, 1))  # 输出层的偏置
18
19    def forward(self, inputs):
20        # 初始化隐藏状态和输出
21        self.h = np.zeros((self.hidden_size, 1))
22        self.outputs = []
23        for x in inputs:
24            # 更新隐藏状态
25            self.h = np.tanh(np.dot(self.Wxh, x) + np.dot(self.Whh, self.h) + self.bh)
26            # 计算输出
27            y = np.dot(self.Why, self.h) + self.by
28            # 应用sigmoid激活函数
29            output = sigmoid(y)
30            self.outputs.append(output)
31        return self.outputs
32
33    def backward(self, inputs, targets, learning_rate=0.1):
34        # 初始化梯度
35        dWxh = np.zeros_like(self.Wxh)
36        dWhh = np.zeros_like(self.Whh)
37        dWhy = np.zeros_like(self.Why)
38        dbh = np.zeros_like(self.bh)
39        dby = np.zeros_like(self.by)
40        dh_next = np.zeros_like(self.h)
41
42        for i in reversed(range(len(inputs))):
43            # 计算输出误差
44            dy = self.outputs[i] - targets[i]
45            # 计算输出层的梯度
46            dWhy += np.dot(dy, self.h.T)
47            dby += dy
48            # 计算隐藏层的误差
49            dh = np.dot(self.Why.T, dy) + dh_next
50            # 应用tanh的导数
51            dh_raw = (1 - self.h ** 2) * dh
52            # 计算隐藏层的梯度
53            dWxh += np.dot(dh_raw, inputs[i].T)
54            dWhh += np.dot(dh_raw, self.h.T)
55            dbh += dh_raw
56            # 更新dh_next
57            dh_next = np.dot(self.Whh.T, dh_raw)
58
59        for dparam in [dWxh, dWhh, dWhy, dbh, dby]:
60            np.clip(dparam, -5, 5, out=dparam)  # 防止梯度爆炸
61
62        # 更新权重和偏置
63        self.Wxh -= learning_rate * dWxh
64        self.Whh -= learning_rate * dWhh
65        self.Why -= learning_rate * dWhy
66        self.bh -= learning_rate * dbh
67        self.by -= learning_rate * dby
68
69
70# 测试代码
71# 定义数据和标签
72inputs = [np.array([[1], [0], [1]]), np.array([[0], [1], [0]])]
73targets = [np.array([[1]]), np.array([[0]])]
74
75input_size = inputs[0].shape[0]
76hidden_size = 25
77output_size = targets[0].shape[0]
78
79# 创建RNN模型，并进行训练
80rnn = RNN(input_size, hidden_size, output_size)
81for epoch in range(1000):
82    outputs = rnn.forward(inputs)
83    loss = np.mean((np.array(outputs)-np.array(targets)) ** 2)
84    rnn.backward(inputs, targets)
85    if (epoch + 1) % 100 == 0:
86        print("次数:", epoch + 1, "误差:", loss)
87
88# 在新数据上进行预测
89new_input = np.array([[1], [1], [1]])
90output = rnn.forward([new_input])
91print("输入:", new_input.flatten())
92print("输出:", output)