难以在 2 层 RNN 中正确设置批量大小

Posted 2023-02-16

【中文标题】难以在 2 层 RNN 中正确设置批量大小

【英文标题】：Difficulty setting batch size correctltly in 2 layer RNN

【发布时间】：2021-11-05 07:11:42

【问题描述】：

我正在构建一个 RNN，它在输出中为 11 个维度生成多类分类输出。输入是我从预训练的手套模型中提取的词嵌入。

我得到的错误是（问题末尾的完整回溯）：

ValueError：预期输入 batch_size (1) 与目标 batch_size (11) 匹配。

请注意，这里我使用batch_size=1，错误提示“预期批处理大小 1 与目标批处理大小 (11) 匹配”。但是，如果我将批量大小更改为 11，错误将更改为：

ValueError：预期输入 batch_size (11) 与目标 batch_size (121) 匹配。

我认为错误来自text 的形状，即torch.Size([11, 300])，它缺少序列长度，但我认为如果我不分配序列长度，它默认为 1。但是，我这样做了不知道怎么加进去。

训练循环：

def train(model, device, train_loader, valid_loader, epochs, learning_rate):

  criterion = nn.CrossEntropyLoss()
  optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
  
  train_loss, validation_loss = [], []
  train_acc, validation_acc = [], []

  for epoch in range(epochs):
    #train
    model.train()
    running_loss = 0.
    correct, total = 0, 0
    steps = 0
    for idx, batch in enumerate(train_loader):
      text = batch["Sample"].to(device)
      target = batch['Class'].to(device)
      print(text.shape, target.shape)
      text, target = text.to(device), target.to(device)
      # add micro for coding training loop
      optimizer.zero_grad()
      print(text.shape)
      output, hidden = model(text.unsqueeze(1))
      #print(output.shape, target.shape, target.view(-1).shape)
      loss = criterion(output, target.view(-1))
      loss.backward()
      optimizer.step()
      steps += 1
      running_loss += loss.item()

      # get accuracy
      _, predicted = torch.max(output, 1)
      print(predicted)
      #predicted = torch.round(output.squeeze())
      total += target.size(0)
      correct += (predicted == target).sum().item()

    train_loss.append(running_loss/len(train_loader))
    train_acc.append(correct/total)

    print(f'Epoch: epoch + 1, '
          f'Training Loss: running_loss/len(train_loader):.4f, '
          f'Training Accuracy: 100*correct/total: .2f%')

    # evaluate on validation data
    model.eval()
    running_loss = 0.
    correct, total = 0, 0

    with torch.no_grad():
      for idx, batch in enumerate(valid_loader):
        text = batch["Sample"].to(device)
        print(type(text), text.shape)
        target = batch['Class'].to(device)
        target = torch.autograd.Variable(target).long()
        text, target = text.to(device), target.to(device)

        optimizer.zero_grad()
        output = model(text)
        
        loss = criterion(output, target)
        running_loss += loss.item()

        # get accuracy
        _, predicted = torch.max(output, 1)
        #predicted = torch.round(output.squeeze())
        total += target.size(0)
        correct += (predicted == target).sum().item()

    validation_loss.append(running_loss/len(valid_loader))
    validation_acc.append(correct/total)

    print (f'Validation Loss: running_loss/len(valid_loader):.4f, '
           f'Validation Accuracy: 100*correct/total: .2f%')

  return train_loss, train_acc, validation_loss, validation_acc

这就是我所说的训练循环：

# Model hyperparamters
#vocab_size = len(word_array)
learning_rate = 1e-3
hidden_dim = 100
output_size = 11
input_size = 300
epochs = 10
n_layers = 2

# Initialize model, training and testing
set_seed(SEED)
vanilla_rnn_model = VanillaRNN(input_size, output_size, hidden_dim, n_layers)
vanilla_rnn_model.to(DEVICE)
vanilla_rnn_start_time = time.time()
vanilla_train_loss, vanilla_train_acc, vanilla_validation_loss, vanilla_validation_acc = train(vanilla_rnn_model,
                                                                                               DEVICE,
                                                                                               train_loader,
                                                                                               valid_loader,
                                                                                               epochs = epochs,
                                                                                               learning_rate = learning_rate)

这就是我创建数据加载器的方式：

# Splitting dataset
# define a batch_size, I'll use 4 as an example
batch_size = 1

train_dset = CustomDataset(X2, y)  # create data set
train_loader = DataLoader(train_dset, batch_size=batch_size, shuffle=True) #load data with batch size
valid_dset = CustomDataset(X2, y)
valid_loader = DataLoader(valid_dset, batch_size=batch_size, shuffle=True)

g_seed = torch.Generator()
g_seed.manual_seed(SEED)

完整的追溯：

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-23-bfd2f8f3456f> in <module>()
     19                                                                                                valid_loader,
     20                                                                                                epochs = epochs,
---> 21                                                                                                learning_rate = learning_rate)
     22 print("--- Time taken to train = %s seconds ---" % (time.time() - vanilla_rnn_start_time))
     23 #test_accuracy = test(vanilla_rnn_model, DEVICE, test_iter)

3 frames
<ipython-input-22-16748701034f> in train(model, device, train_loader, valid_loader, epochs, learning_rate)
     47       output, hidden = model(text.unsqueeze(1))
     48       #print(output.shape, target.shape, target.view(-1).shape)
---> 49       loss = criterion(output, target.view(-1))
     50       loss.backward()
     51       optimizer.step()

/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
   1049         if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
   1050                 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1051             return forward_call(*input, **kwargs)
   1052         # Do not call functions when jit is used
   1053         full_backward_hooks, non_full_backward_hooks = [], []

/usr/local/lib/python3.7/dist-packages/torch/nn/modules/loss.py in forward(self, input, target)
   1119     def forward(self, input: Tensor, target: Tensor) -> Tensor:
   1120         return F.cross_entropy(input, target, weight=self.weight,
-> 1121                                ignore_index=self.ignore_index, reduction=self.reduction)
   1122 
   1123 

/usr/local/lib/python3.7/dist-packages/torch/nn/functional.py in cross_entropy(input, target, weight, size_average, ignore_index, reduce, reduction)
   2822     if size_average is not None or reduce is not None:
   2823         reduction = _Reduction.legacy_get_string(size_average, reduce)
-> 2824     return torch._C._nn.cross_entropy_loss(input, target, weight, _Reduction.get_enum(reduction), ignore_index)
   2825 
   2826 

ValueError: Expected input batch_size (1) to match target batch_size (11).

【参考方案1】：

您不应该使用.view(-1)。这一行：

loss = criterion(output, target.view(-1))

应该是：

loss = criterion(output, target)

它有效地删除了您的批次维度。对于batch_size=1，它将(1, 11) 更改为(11,)。当您将batch_size 更改为11 时，视图将形状从(11, 11) 更改为(121,)，因此出现错误。

【讨论】：

成功了，我会在几分钟内更新到正确的解决方案。不幸的是，我现在收到一个新错误``` RuntimeError: multi-target not supported at /pytorch/aten/src/THCUNN/generic/ClassNLLCriterion.cu:15 ``` 有什么想法吗？如果您愿意，我可以为积分创建一个新帖子。

@Aswiderski 我们不是为了要点...别担心...只有在您确定它不是重复的并且可以为人们提供足够信息的情况下才打开一个问题帮助。您现在遇到的问题已回答here。如果您打开一个问题，它将作为重复关闭。简而言之：PyTorch 期望目标为整数，而您提供的是 one-hot 张量。