Python PyTorch多层感知机实战:早停训练与手写数字识别

from copy import deepcopyimport matplotlib.pyplot as pltimport numpy as npimport torchfrom sklearn.datasets import load_digitsfrom sklearn.metrics import accuracy_score, classification_report, confusion_matrixfrom sklearn.model_selection import train_test_splitfrom torch import nnfrom torch.utils.data import DataLoader, TensorDataset# 固定NumPy、PyTorch和CPU线程设置，减少重复运行差异SEED = 20260821np.random.seed(SEED)torch.manual_seed(SEED)torch.set_num_threads(1)torch.use_deterministic_algorithms(True)device = torch.device(”cpu”)# 使用无需联网下载的手写数字数据，并把像素缩放到0至1digits = load_digits()X = (digits.data / 16.0).astype(”float32”)y = digits.target.astype(”int64”)# 先留出测试集，再从开发数据中划出验证集X_dev, X_test, y_dev, y_test = train_test_split(    X, y, test_size=0.15, stratify=y, random_state=SEED)X_train, X_val, y_train, y_val = train_test_split(    X_dev, y_dev, test_size=0.1768, stratify=y_dev, random_state=SEED)# 转成PyTorch张量，只有训练批次需要随机打乱train_ds = TensorDataset(torch.from_numpy(X_train), torch.from_numpy(y_train))val_ds = TensorDataset(torch.from_numpy(X_val), torch.from_numpy(y_val))test_ds = TensorDataset(torch.from_numpy(X_test), torch.from_numpy(y_test))# 单独固定DataLoader生成器，使批次顺序可复现loader_generator = torch.Generator().manual_seed(SEED)train_loader = DataLoader(    train_ds, batch_size=64, shuffle=True, generator=loader_generator)val_loader = DataLoader(val_ds, batch_size=256, shuffle=False)test_loader = DataLoader(test_ds, batch_size=256, shuffle=False)print(f”全部样本：{len(X)}，类别数：{len(np.unique(y))}”)print(f”训练/验证/测试：{len(X_train)}/{len(X_val)}/{len(X_test)}”)print(f”特征维度：{X.shape[1]}，图像尺寸：{digits.images.shape[1:]}”)print(f”缩放后像素范围：{X.min():.1f} 至 {X.max():.1f}”)

全部样本：1797，类别数：10训练/验证/测试：1257/270/270特征维度：64，图像尺寸：(8, 8)缩放后像素范围：0.0 至 1.0

# 两个隐藏层负责学习像素之间的非线性组合model = nn.Sequential(    nn.Linear(64, 128),    nn.ReLU(),    nn.Dropout(0.20),    nn.Linear(128, 64),    nn.ReLU(),    nn.Linear(64, 10)).to(device)# 交叉熵接收原始logits，Adam负责更新全部参数loss_fn = nn.CrossEntropyLoss()optimizer = torch.optim.Adam(    model.parameters(), lr=0.001, weight_decay=1e-4)# 同一函数通过training参数切换训练模式和评估模式def run_epoch(data_loader, training):    model.train(training)    loss_sum = 0.0    correct = 0    count = 0    context = torch.enable_grad() if training else torch.no_grad()    with context:        for xb, yb in data_loader:            xb, yb = xb.to(device), yb.to(device)            if training:                optimizer.zero_grad()            logits = model(xb)            loss = loss_fn(logits, yb)            if training:                loss.backward()                optimizer.step()            loss_sum += loss.item() * len(yb)            correct += (logits.argmax(dim=1) == yb).sum().item()            count += len(yb)    return loss_sum / count, correct / count# 记录每轮指标，并等待最多12轮无改善history = []best_loss = float(”inf”)best_epoch = 0best_state = Nonepatience = 12wait = 0for epoch in range(1, 121):    train_loss, train_acc = run_epoch(train_loader, training=True)    val_loss, val_acc = run_epoch(val_loader, training=False)    history.append({        ”epoch”: epoch,        ”train_loss”: train_loss,        ”val_loss”: val_loss,        ”train_acc”: train_acc,        ”val_acc”: val_acc    })# 验证损失改善时深拷贝权重，避免后续训练覆盖最佳状态    if val_loss < best_loss - 1e-4:        best_loss = val_loss        best_epoch = epoch        best_state = deepcopy(model.state_dict())        wait = 0    else:        wait += 1# 连续达到耐心轮数后结束，不再消耗额外训练轮次    if wait >= patience:        breakstopped_epoch = history[-1][”epoch”]last_state = deepcopy(model.state_dict())best_row = history[best_epoch - 1]param_count = sum(p.numel() for p in model.parameters())print(f”可训练参数：{param_count}”)print(f”停止轮次：{stopped_epoch}”)print(f”最佳轮次：{best_epoch}”)print(f”最低验证损失：{best_loss:.4f}”)print(f”最佳轮训练/验证准确率：{best_row['train_acc']:.4f}/{best_row['val_acc']:.4f}”)

可训练参数：17226停止轮次：39最佳轮次：27最低验证损失：0.1167最佳轮训练/验证准确率：0.9809/0.9593