图片详情
—
Skip to content

check_gradients.py ​

返回章节 · 配套代码说明

复制以下内容并保存为 check_gradients.py,使用方法见配套代码说明。

python
import random
from nn import MLP, check_gradient

configs = [  # (隐藏激活, 输出激活, 损失, 结构, 随机种子)
    ("tanh",    "linear",  "mse", [3, 4, 2],    3),
    ("relu",    "linear",  "mse", [2, 5, 1],    7),
    ("sigmoid", "sigmoid", "bce", [2, 3, 1],    11),
    ("tanh",    "linear",  "mse", [4, 3, 3, 2], 5),
]
for hidden, out, loss, sizes, seed in configs:
    net = MLP(sizes, hidden_activation=hidden, output_activation=out,
              loss=loss, seed=seed)
    rng = random.Random(seed)
    X = [[rng.uniform(-1, 1) for _ in range(sizes[0])] for _ in range(5)]
    if loss == "bce":
        Y = [[float(rng.random() > 0.5) for _ in range(sizes[-1])] for _ in range(5)]
    else:
        Y = [[rng.uniform(-1, 1) for _ in range(sizes[-1])] for _ in range(5)]
    err = check_gradient(net, X, Y)
    print(f"{hidden:>7} + {loss.upper():<3} {str(sizes):<14} "
          f"参数量 {net.n_params():>3}   最大相对误差 {err:.2e}")

《原子智能》· 纸质出版预备版 · PolyAI Team 著