Day 36 MLP神经网络的训练

Day 36 · MLP神经网络的训练

pytorch和cuda的安装有很多教程，这里就不多赘述了。

importtorch torch.cuda

<module 'torch.cuda' from '/home/ubuntu24/anaconda3/envs/torch-gpu/lib/python3.13/site-packages/torch/cuda/__init__.py'>

importtorch# 检查CUDA是否可用iftorch.cuda.is_available():print("CUDA可用！")# 获取可用的CUDA设备数量device_count=torch.cuda.device_count()print(f"可用的CUDA设备数量:{device_count}")# 获取当前使用的CUDA设备索引current_device=torch.cuda.current_device()print(f"当前使用的CUDA设备索引:{current_device}")# 获取当前CUDA设备的名称device_name=torch.cuda.get_device_name(current_device)print(f"当前CUDA设备的名称:{device_name}")# 获取CUDA版本cuda_version=torch.version.cudaprint(f"CUDA版本:{cuda_version}")else:print("CUDA不可用。")

CUDA可用！ 可用的CUDA设备数量: 1 当前使用的CUDA设备索引: 0 当前CUDA设备的名称: NVIDIA GeForce RTX 4070 Laptop GPU CUDA版本: 12.4

数据的准备

# 导入3分类的鸢尾花数据集fromsklearn.datasetsimportload_irisfromsklearn.model_selectionimporttrain_test_splitimportnumpyasnp iris=load_iris()X=iris.data y=iris.target X_train,X_test,y_train,y_test=train_test_split(X,y,test_size=0.2,random_state=42)print(X_train.shape)print(y_train.shape)print(X_test.shape)print(y_test.shape)

(120, 4) (120,) (30, 4) (30,)

# 神经网络对于输入数据敏感，因此要对输入的数据进行归一化处理fromsklearn.preprocessingimportMinMaxScaler scaler=MinMaxScaler()X_train=scaler.fit_transform(X_train)X_test=scaler.transform(X_test)

# 将数据转换为张量，Pytorch使用张量进行训练，张量可以理解为特殊的数组X_train=torch.FloatTensor(X_train)y_train=torch.LongTensor(y_train)X_test=torch.FloatTensor(X_test)y_test=torch.LongTensor(y_test)

模型设计

importtorchimporttorch.nnasnnimporttorch.optimasoptim# 定义MLP模型model=nn.Sequential(nn.Linear(4,10),nn.ReLU(),nn.Linear(10,3))# 或者classMLP(nn.Module):def__init__(self):super().__init__()self.fc1=nn.Linear(4,10)# 输入层到隐藏层self.relu=nn.ReLU()# 引入非线性self.fc2=nn.Linear(10,3)# 隐藏层到输出层# 输出层不需要激活函数，因为后面会用到交叉熵函数cross_entropy，交叉熵函数内部有softmax函数，会把输出转化为概率defforward(self,x):out=self.fc1(x)out=self.relu(out)out=self.fc2(out)returnout# MLP_model=MLP()

# 分类问题使用交叉熵损失函数criterion=nn.CrossEntropyLoss()# 使用Adam优化器optimizer=optim.Adam(model.parameters(),lr=0.01)

训练

采用交叉熵损失 + Adam 优化器。训练前先把模型和数据移动到同一设备，随后在循环中维护损失、准确率列表。

device=torch.device('cuda'iftorch.cuda.is_available()else'cpu')model=model.to(device)X_train=X_train.to(device)y_train=y_train.to(device)X_test=X_test.to(device)y_test=y_test.to(device)num_epochs=20_000 train_losses,val_losses=[],[]train_accuracies,val_accuracies=[],[]# 计算准确率defcalculate_accuracy(logits,labels):preds=torch.argmax(logits.detach(),dim=1)return(preds==labels).float().mean().item()forepochinrange(1,num_epochs+1):model.train()optimizer.zero_grad()outputs=model(X_train)train_loss=criterion(outputs,y_train)train_loss.backward()optimizer.step()train_losses.append(train_loss.item())train_accuracies.append(calculate_accuracy(outputs,y_train))model.eval()withtorch.no_grad():val_outputs=model(X_test)val_loss=criterion(val_outputs,y_test).item()val_acc=calculate_accuracy(val_outputs,y_test)val_losses.append(val_loss)val_accuracies.append(val_acc)ifepoch%400==0:print(f'Epoch [{epoch}/{num_epochs}] 'f'train_loss={train_loss.item():.4f}val_loss={val_loss:.4f}'f'train_acc={train_accuracies[-1]:.4f}val_acc={val_acc:.4f}')

Epoch [400/20000] train_loss=0.0629 val_loss=0.0538 train_acc=0.9750 val_acc=0.9667 Epoch [800/20000] train_loss=0.0497 val_loss=0.0292 train_acc=0.9833 val_acc=1.0000 Epoch [1200/20000] train_loss=0.0473 val_loss=0.0203 train_acc=0.9833 val_acc=1.0000 Epoch [1600/20000] train_loss=0.0468 val_loss=0.0173 train_acc=0.9833 val_acc=1.0000 Epoch [2000/20000] train_loss=0.0467 val_loss=0.0161 train_acc=0.9833 val_acc=1.0000 Epoch [2400/20000] train_loss=0.0466 val_loss=0.0157 train_acc=0.9833 val_acc=1.0000 Epoch [2800/20000] train_loss=0.0466 val_loss=0.0156 train_acc=0.9833 val_acc=1.0000 Epoch [3200/20000] train_loss=0.0466 val_loss=0.0155 train_acc=0.9833 val_acc=1.0000 Epoch [3600/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [4000/20000] train_loss=0.0466 val_loss=0.0154 train_acc=0.9833 val_acc=1.0000 Epoch [4400/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [4800/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [5200/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [5600/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [6000/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [6400/20000] train_loss=0.0466 val_loss=0.0154 train_acc=0.9833 val_acc=1.0000 Epoch [6800/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [7200/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [7600/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [8000/20000] train_loss=0.0466 val_loss=0.0154 train_acc=0.9833 val_acc=1.0000 Epoch [8400/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [8800/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [9200/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [9600/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [10000/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [10400/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [10800/20000] train_loss=0.0466 val_loss=0.0152 train_acc=0.9833 val_acc=1.0000 Epoch [11200/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [11600/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [12000/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [12400/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [12800/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [13200/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [13600/20000] train_loss=0.0466 val_loss=0.0150 train_acc=0.9833 val_acc=1.0000 Epoch [14000/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [14400/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [14800/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [15200/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [15600/20000] train_loss=0.0466 val_loss=0.0152 train_acc=0.9833 val_acc=1.0000 Epoch [16000/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [16400/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [16800/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [17200/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [17600/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [18000/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [18400/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [18800/20000] train_loss=0.0466 val_loss=0.0155 train_acc=0.9833 val_acc=1.0000 Epoch [19200/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000 Epoch [19600/20000] train_loss=0.0466 val_loss=0.0152 train_acc=0.9833 val_acc=1.0000 Epoch [20000/20000] train_loss=0.0466 val_loss=0.0153 train_acc=0.9833 val_acc=1.0000

可视化

有了损失/准确率数组，画双子图就能一眼看出模型是否过拟合或欠拟合。实践中建议在这里记录实验备注，比如 epoch 数、学习率、是否使用 GPU，方便未来对比。

importmatplotlib.pyplotasplt epochs=range(1,num_epochs+1)plt.figure(figsize=(12,5))plt.subplot(1,2,1)plt.plot(epochs,train_losses,label='Train Loss')plt.plot(epochs,val_losses,label='Validation Loss')plt.xlabel('Epoch')plt.ylabel('Loss')plt.title('Loss over Epochs')plt.legend()plt.subplot(1,2,2)plt.plot(epochs,train_accuracies,label='Train Accuracy')plt.plot(epochs,val_accuracies,label='Validation Accuracy')plt.xlabel('Epoch')plt.ylabel('Accuracy')plt.title('Accuracy over Epochs')plt.legend()plt.tight_layout()plt.show()

@浙大疏锦行