设备类型 cuda 的预期对象,但在 Pytorch 中获得了设备类型 cpu

Posted

技术标签:

【中文标题】设备类型 cuda 的预期对象,但在 Pytorch 中获得了设备类型 cpu【英文标题】:Expected object of device type cuda but got device type cpu in Pytorch 【发布时间】:2020-03-07 00:43:06 【问题描述】:

我有以下计算损失函数的代码:

class MSE_loss(nn.Module):
    """ 
    : metric: L1, L2 norms or cosine similarity
    : mode: training or evaluation mode
    """

    def __init__(self,metric, mode, weighted_sum = False):
        super(MSE_loss, self).__init__()
        self.metric = metric.lower()
        self.loss_function = nn.MSELoss()
        self.mode = mode.lower()
        self.weighted_sum = weighted_sum

    def forward(self, output1, output2, labels):
        self.labels = labels         
        self.linear = nn.Linear(output1.size()[0],1)

        if self.metric == 'cos':
            self.d= F.cosine_similarity(output1, output2)
        elif self.metric == 'l1':
            self.d = torch.abs(output1-output2)
        elif self.metric == 'l2':
            self.d = torch.sqrt((output1-output2)**2)

        def dimensional_reduction(forward):
            if self.weighted_sum:
                distance = self.linear(self.d)
            else:
                distance = torch.mean(self.d,1)
            return distance

        def estimate_loss(forward):
            distance = dimensional_reduction(self.d)
            pred = torch.exp(-distance)
            pred = torch.round(pred)
            loss = self.loss_function(pred, self.labels)
            return pred, loss

        pred, loss = estimate_loss(self.d)

        if self.mode == 'training':
            return loss
        else:
            return pred, loss

给定

criterion = MSE_loss('l1','training', weighted_sum = True)

我想在实现标准时通过 self.linear 神经元后得到距离。但是,我收到错误提示“设备类型为 cuda 的预期对象,但在调用 _th_addmm 时获得了参数 #1 'self' 的设备类型 cpu”,表明出现问题。我省略了代码的第一部分,但我提供了整个错误消息,以便您了解发生了什么。

RuntimeError                              Traceback (most recent call last)
<ipython-input-253-781ed4791260> in <module>()
      7 criterion = MSE_loss('l1','training', weighted_sum = True)
      8 
----> 9 train(test_net, train_loader, 10, batch_size, optimiser, clip, criterion)

<ipython-input-207-02fecbfe3b1c> in train(SNN, dataloader, epochs, batch_size, optimiser, clip, criterion)
     57 
     58             # calculate the loss and perform backprop
---> 59             loss = criterion(output1, output2, labels)
     60             a = [[n,p, p.grad] for n,p in SNN.named_parameters()]
     61 

~/.conda/envs/dalkeCourse/lib/python3.6/site-packages/torch/nn/modules/module.py in __call__(self, *input, **kwargs)
    539             result = self._slow_forward(*input, **kwargs)
    540         else:
--> 541             result = self.forward(*input, **kwargs)
    542         for hook in self._forward_hooks.values():
    543             hook_result = hook(self, input, result)

<ipython-input-248-fb88b987ce71> in forward(self, output1, output2, labels)
     49             return pred, loss
     50 
---> 51         pred, loss = estimate_loss(self.d)
     52 
     53         if self.mode == 'training':

<ipython-input-248-fb88b987ce71> in estimate_loss(forward)
     43 
     44         def estimate_loss(forward):
---> 45             distance = dimensional_reduction(self.d)
     46             pred = torch.exp(-distance)
     47             pred = torch.round(pred)

<ipython-input-248-fb88b987ce71> in dimensional_reduction(forward)
     36             else:
     37                 if self.weighted_sum:
---> 38                     self.d = self.linear(self.d)
     39                 else:
     40                     self.d = torch.mean(self.d,1)

~/.conda/envs/dalkeCourse/lib/python3.6/site-packages/torch/nn/modules/module.py in __call__(self, *input, **kwargs)
    539             result = self._slow_forward(*input, **kwargs)
    540         else:
--> 541             result = self.forward(*input, **kwargs)
    542         for hook in self._forward_hooks.values():
    543             hook_result = hook(self, input, result)

~/.conda/envs/dalkeCourse/lib/python3.6/site-packages/torch/nn/modules/linear.py in forward(self, input)
     85 
     86     def forward(self, input):
---> 87         return F.linear(input, self.weight, self.bias)
     88 
     89     def extra_repr(self):

~/.conda/envs/dalkeCourse/lib/python3.6/site-packages/torch/nn/functional.py in linear(input, weight, bias)
   1368     if input.dim() == 2 and bias is not None:
   1369         # fused op is marginally faster
-> 1370         ret = torch.addmm(bias, input, weight.t())
   1371     else:
   1372         output = input.matmul(weight.t())

RuntimeError: Expected object of device type cuda but got device type cpu for argument #1 'self' in call to _th_addmm

self.d虽然是张量,但是已经传入GPU了,如下图:

self.d =
tensor([[3.7307e-04, 8.4476e-04, 4.0426e-04,  ..., 4.2015e-04, 1.7830e-04,
         1.2833e-04],
        [3.9271e-04, 4.8325e-04, 9.5238e-04,  ..., 1.5126e-04, 1.3420e-04,
         3.9260e-04],
        [1.9278e-04, 2.6530e-04, 8.6903e-04,  ..., 1.6985e-05, 9.5103e-05,
         1.9610e-04],
        ...,
        [1.8257e-05, 3.1304e-04, 4.6398e-04,  ..., 2.7327e-04, 1.1909e-04,
         1.5069e-04],
        [1.7577e-04, 3.4820e-05, 9.4168e-04,  ..., 3.2848e-04, 2.2514e-04,
         5.4275e-05],
        [4.2916e-04, 1.6155e-04, 9.3186e-04,  ..., 1.0950e-04, 2.5083e-04,
         3.7374e-06]], device='cuda:0', grad_fn=<AbsBackward>)

【问题讨论】:

【参考方案1】:

我有同样的问题,结果我应该使用 customized_block = nn.ModuleList([]) 而不是 customized_block = [] 定义模型时。

由于普通列表中的模块不会被识别为nn.Module,因此在调用model.cuda()时不会被放到GPU上。

【讨论】:

【参考方案2】:

作为补充或者一般性的回答,每次遇到这种cudacpu不匹配的错误,首先要检查以下三点:

    您是否将model 放在cuda 上,换句话说,您是否有类似的代码:model = nn.DataParallel(model, device_ids=None).cuda() 是否将input data 放在cuda 上,比如input_data.cuda() 您是否将tensor 放在cuda 上,例如:loss_sum = torch.tensor([losses.sum], dtype=torch.float32, device=device)

Emm,如果你做到了这三个检查,也许你会解决你的问题,祝你好运。

【讨论】:

【参考方案3】:

我在构建我的模型时也遇到了同样的问题,最后我发现这是因为我重新训练了我的模型的全连接层,像这样:

net.to(device)
pre_trained_model=model_path
missing_keys,unexpected_keys=net.load_state_dict(torch.load(pre_trained_model),strict=False)
net.fc=nn.Linear(inchannel,CLASSES)

虽然模型是传输到 cuda 的,但重新定义的 fc 不是,所以最后一行应该是:

net.fc=nn.Linear(inchannel,CLASSES).to(device)

所以检查一下这种情况是否会有所帮助。

【讨论】:

net.fc 是什么?你能解释一下为什么netto(device)net.fc 不是吗?【参考方案4】:

MSE_lossforward 中,您定义了一个可能仍在CPU 中的线性层(您没有提供MCVE,所以我只能假设) :

self.linear = nn.Linear(output1.size()[0], 1)

如果你想试试看是不是这个问题,你可以:

self.linear = nn.Linear(output1.size()[0], 1).cuda()

但是,如果self.d 在 CPU 中,那么它将再次失败。为了解决这个问题,您可以通过执行以下操作将线性移动到 self.d 张量的同一设备:

def forward(self, output1, output2, labels):
    self.labels = labels         
    self.linear = nn.Linear(output1.size()[0], 1)

    if self.metric == 'cos':
        self.d = F.cosine_similarity(output1, output2)
    elif self.metric == 'l1':
        self.d = torch.abs(output1-output2)
    elif self.metric == 'l2':
        self.d = torch.sqrt((output1-output2)**2)

    # move self.linear to the correct device
    self.linear = self.linear.to(self.d.device)

【讨论】:

以上是关于设备类型 cuda 的预期对象,但在 Pytorch 中获得了设备类型 cpu的主要内容,如果未能解决你的问题,请参考以下文章

我更改了标量类型 float 的预期对象,但在 Pytorch 中仍然得到 Long

RuntimeError:预期的标量类型为 Double 的对象,但在调用 _th_max 时为参数 #2 'other' 获得了标量类型 Float

PyTorch - RuntimeError:后端 CPU 的预期对象,但为参数 #2 'weight' 获得了后端 CUDA

为啥我的简单 pytorch 网络不能在 GPU 设备上运行?

无法为cufftComplex数据类型分配CUDA设备内存

有关kaggle中更新cuda与pytorch的版本,并安装pyg