RuntimeError:/pytorch/aten/src/THCUNN/generic/ClassNLLCriterion.cu:15____ 不支持多目标

Posted

技术标签:

【中文标题】RuntimeError:/pytorch/aten/src/THCUNN/generic/ClassNLLCriterion.cu:15____ 不支持多目标【英文标题】:RuntimeError: multi-target not supported at /pytorch/aten/src/THCUNN/generic/ClassNLLCriterion.cu:15____ 【发布时间】:2021-05-03 17:37:33 【问题描述】:

我面临这个错误

我的输入是340的二进制向量,目标是8的二进制向量,对于'" loss = criterion(outputs, stat_batch),我得到了outputs.shape= [64,8] 和stat_batch.shape=[64,8]

这是模型

class MMP(nn.Module):

    def __init__(self, M=1):
        super(MMP, self).__init__()
        # input layer
        self.layer1 = nn.Sequential(
            nn.Conv1d(340, 256,  kernel_size=1, stride=1, padding=0),
            nn.ReLU())
        self.layer2 = nn.Sequential(
            nn.Conv1d(256, 128, kernel_size=1, stride=1, padding=0),
            nn.ReLU())
        self.layer3 = nn.Sequential(
            nn.Conv1d(128, 64, kernel_size=1, stride=1, padding=0),
            nn.ReLU())
        self.drop1 = nn.Sequential(nn.Dropout())
        self.batch1 = nn.BatchNorm1d(128)
        # LSTM
        self.lstm1=nn.Sequential(nn.LSTM(
        input_size=64,
        hidden_size=128,
        num_layers=2,
        bidirectional=True,
        batch_first= True))
        self.fc1 = nn.Linear(128*2,8)
        self.sof = nn.Softmax(dim=-1)

    def forward(self, x):
        out = self.layer1(x)
        out = self.layer2(out)
        out = self.layer3(out)
        out = self.drop1(out)
        out = out.squeeze()
        out = out.unsqueeze(0)
        #out = out.batch1(out)
        out,_ = self.lstm1(out)
        print("lstm",out.shape)
        out = self.fc1(out)
        out =out.squeeze()
        #out = out.squeeze()
        out = self.sof(out)
        return out

#traiin_model
criterion = nn.CrossEntropyLoss()
if CUDA:
    criterion = criterion.cuda()
optimizer = optim.SGD(model.parameters(), lr=LEARNING_RATE, momentum=0.9)

for epoch in range(N_EPOCHES):
    tot_loss=0
    # Training
    for i, (seq_batch, stat_batch) in enumerate(training_generator):
        # Transfer to GPU
        seq_batch, stat_batch = seq_batch.to(device), stat_batch.to(device)
        print(i)
        print(seq_batch)
        print(stat_batch)
        optimizer.zero_grad()
        # Model computation
        seq_batch = seq_batch.unsqueeze(-1)
        outputs = model(seq_batch)
        if CUDA:
            loss = criterion(outputs, stat_batch).float().cuda()
        else:
            loss = criterion(outputs.view(-1), stat_batch.view(-1))
        print(f"Epoch: epoch,number: i, loss:loss.item()...\n\n")

        tot_loss += loss.item(print(f"Epoch: epoch,file_number: i, loss:loss.item()...\n\n"))
        loss.backward()
        optimizer.step()

【问题讨论】:

【参考方案1】:

您的目标stat_batch 的形状必须为(64,),因为nn.CrossEntropyLoss 接受类索引,不是单热编码。

要么适当地构造你的标签张量,要么改用stat_batch.argmax(axis=1)

【讨论】:

标签已经是长张量,我尝试了 stat_batch.argmax(axis=1) 但给了我一个错误,说 ''' loss = criteria(outputs, stat_batch.argmax(axis=1)) TypeError : argmax() 得到了一个意外的关键字参数 'axis'''' “标签已经是长张量” 没关系,它需要是一维的,而不是多维的。您必须运行旧版本的 PyTorch:尝试 stat_batch.argmax(dim=1),或者只是 stat_batch.argmax(1) 非常感谢。是的,它有效,但它仍然具有形状(64,8)。我在 1.99 附近也有很高的损失。你能告诉我绿色复选标记在哪里吗?

以上是关于RuntimeError:/pytorch/aten/src/THCUNN/generic/ClassNLLCriterion.cu:15____ 不支持多目标的主要内容,如果未能解决你的问题,请参考以下文章

RuntimeError: tf.placeholder() 与急切执行不兼容

RuntimeError:预期的标量类型 Double 但发现 Float

RuntimeError:创建新内容类型时出错

multiprocessing进程开发RuntimeError

Python Asyncio - RuntimeError:无法关闭正在运行的事件循环

第二次运行 pyqtgraph 时发生 RuntimeError