训练正常但测试Epoch报错：'int'对象无'double'属性求助

训练Epoch可正常运行，但执行测试Epoch时出现以下错误：

Error: AttributeError: 'int' object has no attribute 'double'
尝试改用.int()方法时，又出现：
Error: AttributeError: 'int' object has no attribute 'int'

相关代码 #

# train loss and accuracy
train_loss = running_loss / len(dataloaders['train'].dataset)
train_acc = running_corrects.double() / len(dataloaders['train'].dataset)

print('{} Loss: {:.4f} Acc: {:.4f}'.format('train', train_loss, train_acc))

writer.add_scalar('Loss/train', train_loss,epoch)
writer.add_scalar('Accuracy/train', train_acc, epoch)

if eval:
  # set the model evaluation mode
  model.eval()

  with torch.no_grad():
    test_correct = 0
    running_loss = 0

    test_iterator = tqdm(enumerate(dataloaders['val']),
                         'Validation',
                         total=len(dataloaders['val']))
    for j, (images,labels) in test_iterator:
      images = images.to(device)
      labels = labels.to(device)

      test_outputs = model(images)

      if (len(criterion)==1):
        loss = criterion[0](outputs, labels)
      elif (len(criterion)==2):
        loss = criterion[0](outputs, labels) + criterion[1](outputs, labels)

      running_loss += loss.item()
      _, test_predicted = torch.max(test_outputs, 1)

      test_correct += torch.sum(test_predicted == labels.data)

  test_loss = running_loss / len(dataloaders['val'].dataset)
  test_acc = test_correct.double() / len(dataloaders['val'].dataset)

  writer.add_scalar('Loss/test', test_loss, epoch)
  writer.add_scalar('Accuracy/test',test_acc, epoch)

  # deep copy the model
  if train_acc > best_acc:
    best_acc = train_acc
    best_model_wts = copy.deepcopy(model.state_dict())

val_acc_history.append(test_acc)
train_accuracy.append(train_acc)
train_losses.append(train_loss)
test_accuracy.append(test_acc)
test_losses.append(test_loss)

错误原因 #

1. 核心问题：test_correct初始化为Python原生int（test_correct = 0），但后续累加的是PyTorch张量（torch.sum(...)返回张量类型）。累加后变量会被强制转换为Python int，调用张量方法.double()自然报错。
2. 次要笔误：测试阶段计算loss时误用了未定义的outputs变量，实际应为模型输出test_outputs，这会导致额外错误。

修复方案 #

1. 初始化张量类型的计数变量：
   将test_correct初始化为同设备的PyTorch张量，和训练阶段的running_corrects保持一致：

   test_correct = torch.tensor(0, device=device)
   

2. 修正loss计算的变量名：
   把outputs替换为test_outputs：

   if (len(criterion)==1):
       loss = criterion[0](test_outputs, labels)
   elif (len(criterion)==2):
       loss = criterion[0](test_outputs, labels) + criterion[1](test_outputs, labels)
   

3. 可选：简化准确率计算：
   无需调用.double()，直接用张量除法即可得到浮点结果：

   test_acc = test_correct / len(dataloaders['val'].dataset)
   

内容的提问来源于stack exchange，提问作者sccomp