{
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "# 实验 C：完整训练一个小网络\n",
        "先猜：训练损失下降是否表示新数据上更准？观察验证曲线，选验证误差最低的参数，最后才计算测试 MAE。比较使用同一训练子集的均值、线性模型与网络。\n",
        "**数据声明：本实验使用固定随机种子生成的模拟教学数据，不是真实材料数据。**\n",
        "先安装依赖，再依次运行下方两格。图片和结果保存在当前目录的 outputs 文件夹。\n"
      ],
      "id": "lab-c-0"
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {},
      "outputs": [],
      "source": [
        "\"\"\"共同的模拟教学数据；不代表真实材料测量或物理规律。\"\"\"\n",
        "from pathlib import Path\n",
        "import csv\n",
        "import numpy as np\n",
        "\n",
        "OUT = Path.cwd() / 'outputs'\n",
        "OUT.mkdir(exist_ok=True)\n",
        "\n",
        "def materials():\n",
        "    rng = np.random.default_rng(42)\n",
        "    temperature = rng.uniform(300, 900, 240)\n",
        "    fraction = rng.uniform(0, 1, 240)\n",
        "    size = rng.uniform(10, 100, 240)\n",
        "    X = np.column_stack([temperature, fraction, size])\n",
        "    y = 2 + .004 * temperature + 3 * fraction**2 - .015 * size + rng.normal(0, .15, 240)\n",
        "    with (OUT / 'simulated_materials.csv').open('w', newline='', encoding='utf-8') as f:\n",
        "        w = csv.writer(f)\n",
        "        w.writerow(['temperature_K', 'fraction', 'size_nm', 'teaching_response'])\n",
        "        w.writerows(np.column_stack([X, y]))\n",
        "    return X, y\n"
      ],
      "id": "lab-c-1"
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {},
      "outputs": [],
      "source": [
        "\"\"\"实验 C：CPU 上完整训练、验证选模、独立测试与参数恢复。\"\"\"\n",
        "import copy\n",
        "import json\n",
        "import numpy as np\n",
        "import torch\n",
        "from torch import nn\n",
        "from torch.utils.data import TensorDataset, DataLoader\n",
        "import matplotlib\n",
        "matplotlib.use('Agg')\n",
        "import matplotlib.pyplot as plt\n",
        "from sklearn.model_selection import train_test_split\n",
        "from sklearn.preprocessing import StandardScaler\n",
        "from sklearn.linear_model import LinearRegression\n",
        "from sklearn.metrics import mean_absolute_error\n",
        "\n",
        "\n",
        "torch.manual_seed(42)\n",
        "torch.set_num_threads(1)\n",
        "X, y = materials()\n",
        "# 与实验 A 相同的最终测试集；只在余下数据里建立验证集。\n",
        "Xa, Xt, ya, yt = train_test_split(X, y, test_size=.25, random_state=42)\n",
        "Xr, Xv, yr, yv = train_test_split(Xa, ya, test_size=.2, random_state=42)\n",
        "sx, sy = StandardScaler().fit(Xr), StandardScaler().fit(yr.reshape(-1, 1))\n",
        "def features(a):\n",
        "    return torch.tensor(sx.transform(a), dtype=torch.float32)\n",
        "def targets(a):\n",
        "    return torch.tensor(sy.transform(a.reshape(-1, 1)), dtype=torch.float32)\n",
        "xr, xv, xt = map(features, (Xr, Xv, Xt))\n",
        "tr, tv = map(targets, (yr, yv))\n",
        "assert tr.shape == (len(yr), 1)\n",
        "loader = DataLoader(TensorDataset(xr, tr), batch_size=32, shuffle=True)\n",
        "model = nn.Sequential(nn.Linear(3, 24), nn.ReLU(), nn.Linear(24, 1))\n",
        "criterion = nn.MSELoss()\n",
        "optimizer = torch.optim.Adam(model.parameters(), lr=.01)\n",
        "best, checkpoint, history = float('inf'), None, []\n",
        "for epoch in range(250):\n",
        "    model.train()\n",
        "    total = 0.\n",
        "    for xb, yb in loader:\n",
        "        optimizer.zero_grad()\n",
        "        loss = criterion(model(xb), yb)\n",
        "        loss.backward()\n",
        "        optimizer.step()\n",
        "        total += loss.item() * len(xb)\n",
        "    model.eval()\n",
        "    with torch.no_grad():\n",
        "        validation = criterion(model(xv), tv).item()\n",
        "    history.append([total / len(xr), validation])\n",
        "    if validation < best:\n",
        "        best, checkpoint = validation, copy.deepcopy(model.state_dict())\n",
        "model.load_state_dict(checkpoint)\n",
        "model.eval()\n",
        "with torch.no_grad():\n",
        "    predicted = sy.inverse_transform(model(xt).numpy()).ravel()\n",
        "# 同时给出同一训练子集上的线性模型及均值基线。\n",
        "baseline = np.full(len(yt), yr.mean())\n",
        "linear = LinearRegression().fit(sx.transform(Xr), yr).predict(sx.transform(Xt))\n",
        "results = {name: float(mean_absolute_error(yt, p)) for name, p in [('mean', baseline), ('linear', linear), ('network', predicted)]}\n",
        "torch.save(model.state_dict(), OUT / 'c_weights.pth')\n",
        "restored = nn.Sequential(nn.Linear(3, 24), nn.ReLU(), nn.Linear(24, 1))\n",
        "restored.load_state_dict(torch.load(OUT / 'c_weights.pth', weights_only=True, map_location='cpu'))\n",
        "restored.eval()\n",
        "with torch.no_grad():\n",
        "    assert torch.allclose(model(xt), restored(xt))\n",
        "# 预测还需配套缩放参数，不能只保存网络权重。\n",
        "np.savez(OUT / 'c_scalers.npz', x_mean=sx.mean_, x_scale=sx.scale_, y_mean=sy.mean_, y_scale=sy.scale_)\n",
        "fig, axes = plt.subplots(1, 2, figsize=(10, 4), constrained_layout=True)\n",
        "axes[0].plot(history, label=['train', 'validation'])\n",
        "axes[0].legend()\n",
        "axes[0].set(xlabel='Epoch', ylabel='MSE (scaled target)', title='Select by validation loss')\n",
        "axes[1].scatter(yt, predicted, s=18)\n",
        "axes[1].plot([min(yt), max(yt)], [min(yt), max(yt)], '--', color='gray')\n",
        "axes[1].set(xlabel='Measured teaching response', ylabel='Prediction', title=f'Network test MAE={results[\"network\"]:.3f}')\n",
        "fig.savefig(OUT / 'c_training.png', dpi=150)\n",
        "plt.close(fig)\n",
        "(OUT / 'c_results.json').write_text(json.dumps(results, indent=2), encoding='utf-8')\n",
        "print('张量形状:', xr.shape, tr.shape)\n",
        "print(json.dumps(results, indent=2))\n"
      ],
      "id": "lab-c-2"
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "## 运行后回答\n",
        "先猜：训练损失下降是否表示新数据上更准？观察验证曲线，选验证误差最低的参数，最后才计算测试 MAE。比较使用同一训练子集的均值、线性模型与网络。\n",
        "提交结果表、输出图、150 字解释和一个局限。改变设置属于新实验；不要看测试结果反复挑设置。\n"
      ],
      "id": "lab-c-3"
    }
  ],
  "metadata": {
    "kernelspec": {
      "display_name": "Python 3",
      "language": "python",
      "name": "python3"
    },
    "language_info": {
      "name": "python",
      "version": "3.12"
    }
  },
  "nbformat": 4,
  "nbformat_minor": 5
}