Appearance
demo_one_step.py
复制以下内容并保存为 demo_one_step.py,使用方法见配套代码说明。
python
"""核对正文中一次完整的前向计算、反向求导与参数更新。"""
from nn import MLP
net = MLP([2, 2, 1], hidden_activation="tanh",
output_activation="linear", loss="mse", seed=0)
net.W = [[[0.8, -0.5], [0.2, 0.9]], [[0.6, -0.7]]]
net.b = [[0.1, -0.2], [0.05]]
x, y = [0.5, -0.3], [0.4]
prediction_before = net.predict(x)[0]
loss_before, dW, db = net.backward(x, y)
# 所有梯度已经用同一组旧参数算好,现在统一更新。
lr = 0.1
for layer in range(len(net.W)):
for neuron in range(len(net.W[layer])):
for source in range(len(net.W[layer][neuron])):
net.W[layer][neuron][source] -= lr * dW[layer][neuron][source]
net.b[layer][neuron] -= lr * db[layer][neuron]
prediction_after = net.predict(x)[0]
loss_after = net.loss_on(x, y)
print(f"预测:{prediction_before:.6f} -> {prediction_after:.6f}")
print(f"损失:{loss_before:.6f} -> {loss_after:.6f}")