线性回归在线训练

在浏览器内用 Pyodide 运行 Python，训练 y = 2x + 1 的线性模型。

参数面板

Pyodide：等待

learning_rate

控制 w 和 b 每轮更新的步长。

0.01

0.0010.08

epochs

训练轮数，越高耗时越长。

noise

样本点围绕真实直线的随机扰动。

0.8

sample_size

生成训练样本的数量。

代码区

import json
import random

params = json.loads(params_json)
learning_rate = float(params["learning_rate"])
epochs = int(params["epochs"])
noise = float(params["noise"])
sample_size = int(params["sample_size"])

random.seed(42)
xs = [i * 10 / max(sample_size - 1, 1) for i in range(sample_size)]
ys = [2 * x + 1 + random.gauss(0, noise) for x in xs]

w = 0.0
b = 0.0
loss_curve = []

for epoch in range(epochs):
    predictions = [w * x + b for x in xs]
    errors = [pred - y for pred, y in zip(predictions, ys)]
    mse = sum(error * error for error in errors) / sample_size
    loss_curve.append({"step": epoch + 1, "loss": mse})

dw = 2 * sum(error * x for error, x in zip(errors, xs)) / sample_size
    db = 2 * sum(errors) / sample_size
    w -= learning_rate * dw
    b -= learning_rate * db

final_predictions = [w * x + b for x in xs]
final_errors = [pred - y for pred, y in zip(final_predictions, ys)]
final_mse = sum(error * error for error in final_errors) / sample_size

result = {
    "w": w,
    "b": b,
    "mse": final_mse,
    "lossCurve": loss_curve,
    "fitData": [
        {"x": x, "y": y, "prediction": pred}
        for x, y, pred in zip(xs, ys, final_predictions)
    ],
}

json.dumps(result)

运行日志

等待运行。

图表结果

调整参数后点击运行，这里会显示轨迹、loss 曲线、指标和结论解释。