简单神经网络反向传播代码

在深度学习中，简单神经网络的反向传播是一种计算梯度的方法，用于优化网络权重，使得预测结果接近真实值。以下是一个简单的用Python和TensorFlow实现的反向传播示例，假设我们有一个只有一个隐藏层的前馈神经网络：

import tensorflow as tf

# 定义模型参数
weights = {
    'hidden': tf.Variable(tf.random.normal([input_size, hidden_units])),
    'output': tf.Variable(tf.random.normal([hidden_units, output_size]))
}
biases = {
    'hidden': tf.Variable(tf.zeros(hidden_units)),
    'output': tf.Variable(tf.zeros(output_size))
}

# 假设输入X，目标y，激活函数用ReLU
def forward_pass(X):
    hidden_layer = tf.nn.relu(tf.add(tf.matmul(X, weights['hidden']), biases['hidden']))
    output = tf.add(tf.matmul(hidden_layer, weights['output']), biases['output'])
    return output

# 模型训练需要的损失函数和优化器
loss_fn = tf.losses.mean_squared_error(y_true, y_pred)
optimizer = tf.optimizers.SGD(learning_rate)

# 反向传播
with tf.GradientTape() as tape:
    y_pred = forward_pass(X)
    loss_value = loss_fn(y_true, y_pred)

gradients = tape.gradient(loss_value, [weights['hidden'], biases['hidden'], weights['output'], biases['output']])
optimizer.apply_gradients(zip(gradients, [weights['hidden'], biases['hidden'], weights['output'], biases['output']]))