文章目录
- Day 39 · 信贷数据集神经网络训练
- 一、数据预处理
- 二、构建 DataLoader 与神经网络
- 三、可视化
- Dropout 模型表现
- 四、小结
Day 39 · 信贷数据集神经网络训练
importpandasaspdimportnumpyasnpimporttorchimporttorch.nnasnnimporttorch.optimasoptimimportmatplotlib.pyplotaspltfromsklearn.model_selectionimporttrain_test_splitimportwarningsimportrandom warnings.filterwarnings('ignore')# 忽略警告信息defset_seed(seed:int=42):random.seed(seed)np.random.seed(seed)torch.manual_seed(seed)torch.cuda.manual_seed_all(seed)torch.backends.cudnn.deterministic=Truetorch.backends.cudnn.benchmark=Falseset_seed(42)data=pd.read_csv('../data.csv')data.head()| Id | Home Ownership | Annual Income | Years in current job | Tax Liens | Number of Open Accounts | Years of Credit History | Maximum Open Credit | Number of Credit Problems | Months since last delinquent | Bankruptcies | Purpose | Term | Current Loan Amount | Current Credit Balance | Monthly Debt | Credit Score | Credit Default | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 0 | 0 | Own Home | 482087.0 | NaN | 0.0 | 11.0 | 26.3 | 685960.0 | 1.0 | NaN | 1.0 | debt consolidation | Short Term | 99999999.0 | 47386.0 | 7914.0 | 749.0 | 0 |
| 1 | 1 | Own Home | 1025487.0 | 10+ years | 0.0 | 15.0 | 15.3 | 1181730.0 | 0.0 | NaN | 0.0 | debt consolidation | Long Term | 264968.0 | 394972.0 | 18373.0 | 737.0 | 1 |
| 2 | 2 | Home Mortgage | 751412.0 | 8 years | 0.0 | 11.0 | 35.0 | 1182434.0 | 0.0 | NaN | 0.0 | debt consolidation | Short Term | 99999999.0 | 308389.0 | 13651.0 | 742.0 | 0 |
| 3 | 3 | Own Home | 805068.0 | 6 years | 0.0 | 8.0 | 22.5 | 147400.0 | 1.0 | NaN | 1.0 | debt consolidation | Short Term | 121396.0 | 95855.0 | 11338.0 | 694.0 | 0 |
| 4 | 4 | Rent | 776264.0 | 8 years | 0.0 | 13.0 | 13.6 | 385836.0 | 1.0 | NaN | 0.0 | debt consolidation | Short Term | 125840.0 | 93309.0 | 7180.0 | 719.0 | 0 |
一、数据预处理
discrete_features=data.select_dtypes(include=['object']).columns.tolist()print(discrete_features)maps={'Home Ownership':{'Own Home':1,'Rent':2,'Have Mortgage':3,'Home Mortgage':4},'Years in current job':{'< 1 year':1,'1 year':2,'2 years':3,'3 years':4,'4 years':5,'5 years':6,'6 years':7,'7 years':8,'8 years':9,'9 years':10,'10+ years':11},'Term':{'Short Term':0,'Long Term':1}}data=data.replace(maps)# Purpose 独热编码，记得需要将bool类型转换为数值data=pd.get_dummies(data,columns=['Purpose'])data2=pd.read_csv('../data.csv')list_diff=data.columns.difference(data2.columns)data[list_diff]=data[list_diff].astype(int)data.rename(columns={'Term':'Long Term'},inplace=True)# 重命名列continuous_features=data.select_dtypes(include=['int64','float64']).columns.tolist()#把筛选出来的列名转换成列表# 连续特征用中位数补全forfeatureincontinuous_features:mode_value=data[feature].mode()[0]#获取该列的众数。data[feature].fillna(mode_value,inplace=True)#用众数填充该列的缺失值，inplace=True表示直接在原数据上修改。X=data.drop(['Credit Default','Id'],axis=1)# 特征，axis=1表示按列删除y=data['Credit Default']# 标签# 按照8:2划分训练集和测试集X_train,X_test,y_train,y_test=train_test_split(X,y,test_size=0.2,random_state=42)# 80%训练集，20%测试集['Home Ownership', 'Years in current job', 'Purpose', 'Term']fromsklearn.preprocessingimportStandardScaler scaler=StandardScaler()X_train=scaler.fit_transform(X_train)X_test=scaler.transform(X_test)print(X_train.shape)print(X_test.shape)print(y_train.shape)print(y_test.shape)(6000, 30) (1500, 30) (6000,) (1500,)二、构建 DataLoader 与神经网络
- 为
train_loader、test_loader设置可选的 pinned memory，方便 GPU 加速。 - 在每个 epoch 结束后，对训练集与测试集分别计算损失和准确率，并保存下来用于可视化。
- 打印定期的监控日志，便于观察收敛情况。
fromtorch.utils.dataimportDataLoader,TensorDataset device=torch.device('cuda:0'iftorch.cuda.is_available()else'cpu')pin_memory=device.type=='cuda'X_train=torch.FloatTensor(X_train)y_train=torch.FloatTensor(y_train.to_numpy()).unsqueeze(1)X_test=torch.FloatTensor(X_test)y_test=torch.FloatTensor(y_test.to_numpy()).unsqueeze(1)print(X_train.shape)print(X_test.shape)print(y_train.shape)print(y_test.shape)print('---------------------------------')train_dataset=TensorDataset(X_train,y_train)test_dataset=TensorDataset(X_test,y_test)train_loader=DataLoader(train_dataset,batch_size=64,shuffle=True,pin_memory=pin_memory)test_loader=DataLoader(test_dataset,batch_size=256,shuffle=False,pin_memory=pin_memory)model=nn.Sequential(nn.Linear(X_train.shape[1],64),nn.ReLU(),nn.Linear(64,32),nn.ReLU(),nn.Linear(32,1)).to(device)criterion=nn.BCEWithLogitsLoss()optimizer=optim.Adam(model.parameters(),lr=1e-3)num_epochs=300train_losses,test_losses=[],[]train_accuracies,test_accuracies=[],[]forepochinrange(1,num_epochs+1):model.train()running_loss=0.0running_correct=0total_train=0forx_batch,y_batchintrain_loader:x_batch=x_batch.to(device,non_blocking=pin_memory)y_batch=y_batch.to(device,non_blocking=pin_memory)optimizer.zero_grad()outputs=model(x_batch)loss=criterion(outputs,y_batch)loss.backward()optimizer.step()running_loss+=loss.item()*x_batch.size(0)preds=(torch.sigmoid(outputs)>0.5).int()running_correct+=(preds==y_batch.int()).sum().item()total_train+=x_batch.size(0)avg_train_loss=running_loss/total_train avg_train_acc=running_correct/total_train train_losses.append(avg_train_loss)train_accuracies.append(avg_train_acc)model.eval()test_loss=0.0test_correct=0total_test=0withtorch.no_grad():forx_batch,y_batchintest_loader:x_batch=x_batch.to(device,non_blocking=pin_memory)y_batch=y_batch.to(device,non_blocking=pin_memory)outputs=model(x_batch)loss=criterion(outputs,y_batch)test_loss+=loss.item()*x_batch.size(0)preds=(torch.sigmoid(outputs)>0.5).int()test_correct+=(preds==y_batch.int()).sum().item()total_test+=x_batch.size(0)avg_test_loss=test_loss/total_test avg_test_acc=test_correct/total_test test_losses.append(avg_test_loss)test_accuracies.append(avg_test_acc)ifepoch%20==0orepoch==1:print(f'Epoch [{epoch:03d}/{num_epochs}] | 'f'Train Loss:{avg_train_loss:.4f}, Train Acc:{avg_train_acc:.4f}| 'f'Test Loss:{avg_test_loss:.4f}, Test Acc:{avg_test_acc:.4f}')print(f'Final Test Accuracy:{test_accuracies[-1]:.4f}')torch.Size([6000, 30]) torch.Size([1500, 30]) torch.Size([6000, 1]) torch.Size([1500, 1]) --------------------------------- Epoch [001/300] | Train Loss: 0.5566, Train Acc: 0.7417 | Test Loss: 0.5053, Test Acc: 0.7673 Epoch [020/300] | Train Loss: 0.4383, Train Acc: 0.7883 | Test Loss: 0.4872, Test Acc: 0.7587 Epoch [040/300] | Train Loss: 0.4183, Train Acc: 0.7990 | Test Loss: 0.4940, Test Acc: 0.7540 Epoch [060/300] | Train Loss: 0.3951, Train Acc: 0.8105 | Test Loss: 0.5158, Test Acc: 0.7500 Epoch [080/300] | Train Loss: 0.3816, Train Acc: 0.8165 | Test Loss: 0.5445, Test Acc: 0.7560 Epoch [100/300] | Train Loss: 0.3640, Train Acc: 0.8263 | Test Loss: 0.5684, Test Acc: 0.7560 Epoch [120/300] | Train Loss: 0.3565, Train Acc: 0.8320 | Test Loss: 0.5776, Test Acc: 0.7220 Epoch [140/300] | Train Loss: 0.3459, Train Acc: 0.8402 | Test Loss: 0.5914, Test Acc: 0.7213 Epoch [160/300] | Train Loss: 0.3373, Train Acc: 0.8432 | Test Loss: 0.6063, Test Acc: 0.7400 Epoch [180/300] | Train Loss: 0.3273, Train Acc: 0.8488 | Test Loss: 0.6210, Test Acc: 0.7427 Epoch [200/300] | Train Loss: 0.3175, Train Acc: 0.8558 | Test Loss: 0.6753, Test Acc: 0.7007 Epoch [220/300] | Train Loss: 0.3103, Train Acc: 0.8628 | Test Loss: 0.6701, Test Acc: 0.7233 Epoch [240/300] | Train Loss: 0.3035, Train Acc: 0.8618 | Test Loss: 0.6862, Test Acc: 0.7060 Epoch [260/300] | Train Loss: 0.2984, Train Acc: 0.8628 | Test Loss: 0.7258, Test Acc: 0.7120 Epoch [280/300] | Train Loss: 0.2891, Train Acc: 0.8730 | Test Loss: 0.7432, Test Acc: 0.7213 Epoch [300/300] | Train Loss: 0.2828, Train Acc: 0.8750 | Test Loss: 0.7428, Test Acc: 0.7220 Final Test Accuracy: 0.7220三、可视化
下图左侧展示损失曲线，右侧展示准确率曲线。
epochs=range(1,num_epochs+1)fig,axes=plt.subplots(1,2,figsize=(14,5))axes[0].plot(epochs,train_losses,label='Train Loss')axes[0].plot(epochs,test_losses,label='Test Loss')axes[0].set_xlabel('Epoch')axes[0].set_ylabel('Loss')axes[0].set_title('Train vs Test Loss')axes[0].legend()axes[1].plot(epochs,train_accuracies,label='Train Acc')axes[1].plot(epochs,test_accuracies,label='Test Acc')axes[1].set_xlabel('Epoch')axes[1].set_ylabel('Accuracy')axes[1].set_title('Train vs Test Accuracy')axes[1].legend()plt.tight_layout()plt.show()
可以看到，模型在测试集上的表现极差。过拟合很严重。接下来，我们来尝试使用Dropout来减缓过拟合。
# 使用 Dropout + L2 正则重新训练，缓解过拟合set_seed(42)# 确保第二次实验也可复现drop_model=nn.Sequential(nn.Linear(X_train.shape[1],128),nn.ReLU(),nn.Dropout(0.3),nn.Linear(128,64),nn.ReLU(),nn.Dropout(0.3),nn.Linear(64,1)).to(device)weight_decay=1e-4drop_optimizer=optim.Adam(drop_model.parameters(),lr=1e-3,weight_decay=weight_decay)drop_epochs=200drop_train_losses,drop_test_losses=[],[]drop_train_accs,drop_test_accs=[],[]forepochinrange(1,drop_epochs+1):drop_model.train()running_loss=0.0running_correct=0total=0forxb,ybintrain_loader:xb=xb.to(device,non_blocking=pin_memory)yb=yb.to(device,non_blocking=pin_memory)drop_optimizer.zero_grad()logits=drop_model(xb)loss=criterion(logits,yb)loss.backward()drop_optimizer.step()running_loss+=loss.item()*xb.size(0)preds=(torch.sigmoid(logits)>0.5).int()running_correct+=(preds==yb.int()).sum().item()total+=xb.size(0)train_loss=running_loss/total train_acc=running_correct/total drop_train_losses.append(train_loss)drop_train_accs.append(train_acc)drop_model.eval()test_loss=0.0test_correct=0total_test=0withtorch.no_grad():forxb,ybintest_loader:xb=xb.to(device,non_blocking=pin_memory)yb=yb.to(device,non_blocking=pin_memory)logits=drop_model(xb)loss=criterion(logits,yb)test_loss+=loss.item()*xb.size(0)preds=(torch.sigmoid(logits)>0.5).int()test_correct+=(preds==yb.int()).sum().item()total_test+=xb.size(0)avg_test_loss=test_loss/total_test avg_test_acc=test_correct/total_test drop_test_losses.append(avg_test_loss)drop_test_accs.append(avg_test_acc)ifepoch%20==0orepoch==1:print(f'Dropout Epoch [{epoch:03d}/{drop_epochs}] | 'f'Train Loss:{train_loss:.4f}, Train Acc:{train_acc:.4f}| 'f'Test Loss:{avg_test_loss:.4f}, Test Acc:{avg_test_acc:.4f}')print(f'Final Test Accuracy with Dropout/L2:{drop_test_accs[-1]:.4f}')Dropout Epoch [001/200] | Train Loss: 0.5564, Train Acc: 0.7473 | Test Loss: 0.4935, Test Acc: 0.7667 Dropout Epoch [020/200] | Train Loss: 0.4507, Train Acc: 0.7858 | Test Loss: 0.4711, Test Acc: 0.7647 Dropout Epoch [040/200] | Train Loss: 0.4392, Train Acc: 0.7872 | Test Loss: 0.4743, Test Acc: 0.7620 Dropout Epoch [060/200] | Train Loss: 0.4284, Train Acc: 0.7978 | Test Loss: 0.4817, Test Acc: 0.7593 Dropout Epoch [080/200] | Train Loss: 0.4182, Train Acc: 0.8008 | Test Loss: 0.4826, Test Acc: 0.7553 Dropout Epoch [100/200] | Train Loss: 0.4133, Train Acc: 0.8065 | Test Loss: 0.4929, Test Acc: 0.7547 Dropout Epoch [120/200] | Train Loss: 0.4085, Train Acc: 0.8067 | Test Loss: 0.4951, Test Acc: 0.7533 Dropout Epoch [140/200] | Train Loss: 0.3974, Train Acc: 0.8135 | Test Loss: 0.5040, Test Acc: 0.7580 Dropout Epoch [160/200] | Train Loss: 0.3925, Train Acc: 0.8173 | Test Loss: 0.5158, Test Acc: 0.7533 Dropout Epoch [180/200] | Train Loss: 0.3872, Train Acc: 0.8178 | Test Loss: 0.5231, Test Acc: 0.7533 Dropout Epoch [200/200] | Train Loss: 0.3828, Train Acc: 0.8220 | Test Loss: 0.5193, Test Acc: 0.7480 Final Test Accuracy with Dropout/L2: 0.7480Dropout 模型表现
再绘制一次损失/准确率曲线，观察正则化后的收敛情况。
drop_epochs_range=range(1,drop_epochs+1)fig,axes=plt.subplots(1,2,figsize=(14,5))axes[0].plot(drop_epochs_range,drop_train_losses,label='Drop Train Loss')axes[0].plot(drop_epochs_range,drop_test_losses,label='Drop Test Loss')axes[0].set_xlabel('Epoch')axes[0].set_ylabel('Loss')axes[0].set_title('Dropout Model Loss')axes[0].legend()axes[1].plot(drop_epochs_range,drop_train_accs,label='Drop Train Acc')axes[1].plot(drop_epochs_range,drop_test_accs,label='Drop Test Acc')axes[1].set_xlabel('Epoch')axes[1].set_ylabel('Accuracy')axes[1].set_title('Dropout Model Accuracy')axes[1].legend()plt.tight_layout()plt.show()
四、小结
- 我天，效果更差了。(ㄒ o ㄒ)
- 第一阶段模型有明显的过拟合迹象，因此又用 Dropout + L2 重新训练，记录了第二组损失/准确率曲线。结果效果更差了(｡ŏ﹏ŏ)。本人才疏学浅，先这样吧，之后再来研究一下是为啥吧。
- 之后继续在正则化、网络深度、学习率调度或早停等方向尝试。
@浙大疏锦行
