"""实验 C：CPU 上完整训练、验证选模、独立测试与参数恢复。"""
import copy
import json
import numpy as np
import torch
from torch import nn
from torch.utils.data import TensorDataset, DataLoader
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LinearRegression
from sklearn.metrics import mean_absolute_error
from common import materials, OUT

torch.manual_seed(42)
torch.set_num_threads(1)
X, y = materials()
# 与实验 A 相同的最终测试集；只在余下数据里建立验证集。
Xa, Xt, ya, yt = train_test_split(X, y, test_size=.25, random_state=42)
Xr, Xv, yr, yv = train_test_split(Xa, ya, test_size=.2, random_state=42)
sx, sy = StandardScaler().fit(Xr), StandardScaler().fit(yr.reshape(-1, 1))
def features(a):
    return torch.tensor(sx.transform(a), dtype=torch.float32)
def targets(a):
    return torch.tensor(sy.transform(a.reshape(-1, 1)), dtype=torch.float32)
xr, xv, xt = map(features, (Xr, Xv, Xt))
tr, tv = map(targets, (yr, yv))
assert tr.shape == (len(yr), 1)
loader = DataLoader(TensorDataset(xr, tr), batch_size=32, shuffle=True)
model = nn.Sequential(nn.Linear(3, 24), nn.ReLU(), nn.Linear(24, 1))
criterion = nn.MSELoss()
optimizer = torch.optim.Adam(model.parameters(), lr=.01)
best, checkpoint, history = float('inf'), None, []
for epoch in range(250):
    model.train()
    total = 0.
    for xb, yb in loader:
        optimizer.zero_grad()
        loss = criterion(model(xb), yb)
        loss.backward()
        optimizer.step()
        total += loss.item() * len(xb)
    model.eval()
    with torch.no_grad():
        validation = criterion(model(xv), tv).item()
    history.append([total / len(xr), validation])
    if validation < best:
        best, checkpoint = validation, copy.deepcopy(model.state_dict())
model.load_state_dict(checkpoint)
model.eval()
with torch.no_grad():
    predicted = sy.inverse_transform(model(xt).numpy()).ravel()
# 同时给出同一训练子集上的线性模型及均值基线。
baseline = np.full(len(yt), yr.mean())
linear = LinearRegression().fit(sx.transform(Xr), yr).predict(sx.transform(Xt))
results = {name: float(mean_absolute_error(yt, p)) for name, p in [('mean', baseline), ('linear', linear), ('network', predicted)]}
torch.save(model.state_dict(), OUT / 'c_weights.pth')
restored = nn.Sequential(nn.Linear(3, 24), nn.ReLU(), nn.Linear(24, 1))
restored.load_state_dict(torch.load(OUT / 'c_weights.pth', weights_only=True, map_location='cpu'))
restored.eval()
with torch.no_grad():
    assert torch.allclose(model(xt), restored(xt))
# 预测还需配套缩放参数，不能只保存网络权重。
np.savez(OUT / 'c_scalers.npz', x_mean=sx.mean_, x_scale=sx.scale_, y_mean=sy.mean_, y_scale=sy.scale_)
fig, axes = plt.subplots(1, 2, figsize=(10, 4), constrained_layout=True)
axes[0].plot(history, label=['train', 'validation'])
axes[0].legend()
axes[0].set(xlabel='Epoch', ylabel='MSE (scaled target)', title='Select by validation loss')
axes[1].scatter(yt, predicted, s=18)
axes[1].plot([min(yt), max(yt)], [min(yt), max(yt)], '--', color='gray')
axes[1].set(xlabel='Measured teaching response', ylabel='Prediction', title=f'Network test MAE={results["network"]:.3f}')
fig.savefig(OUT / 'c_training.png', dpi=150)
plt.close(fig)
(OUT / 'c_results.json').write_text(json.dumps(results, indent=2), encoding='utf-8')
print('张量形状:', xr.shape, tr.shape)
print(json.dumps(results, indent=2))
