Pytorch CNN 训练中的“RuntimeError：预期标量类型为 Double 但发现 Float”

如何解决Pytorch CNN 训练中的“RuntimeError：预期标量类型为 Double 但发现 Float”

我刚刚开始学习 Pytorch 并创建了我的第一个 CNN。数据集包含 3360 张 RGB 图像，我将它们转换为 [3360,3,224,224] 张量。数据和标签在 dataset(torch.utils.data.TensorDataset) 中。下面是训练代码。

    def train_net():
        dataset = ld.load()
        data_iter = Data.DataLoader(dataset,batch_size=168,shuffle=True)
        net = model.VGG_19()
        summary(net,(3,224),device="cpu")
        loss_func = nn.CrossEntropyLoss()
        optimizer = optim.SGD(net.parameters(),lr=0.001,momentum=0.9,dampening=0.1)
        scheduler = torch.optim.lr_scheduler.StepLR(optimizer,step_size=50,gamma=0.1)
        for epoch in range(5):
            print("epoch:",epoch + 1)
            train_loss = 0
            for i,data in enumerate(data_iter,0):
                x,y = data
                print(x.dtype)
                optimizer.zero_grad()
                out = net(x)
                loss = loss_func(out,y)
                loss.backward()
                optimizer.step()
                train_loss += loss.item()
                if i % 100 == 99:
                    print("loss:",train_loss / 100)
                    train_loss = 0.0
        print("finish train")

然后我有这个错误：

    Traceback (most recent call last):
              File "D:/python/DeepLearning/VGG/train.py",line 52,in <module>
                train_net()
              File "D:/python/DeepLearning/VGG/train.py",line 29,in train_net
                out = net(x)
              File "D:\python\lib\site-packages\torch\nn\modules\module.py",line 727,in _call_impl
                result = self.forward(*input,**kwargs)
              File "D:\python\DeepLearning\VGG\model.py",line 37,in forward
                out = self.conv3_64(x)
              File "D:\python\lib\site-packages\torch\nn\modules\module.py",**kwargs)
              File "D:\python\lib\site-packages\torch\nn\modules\container.py",line 117,in forward
                input = module(input)
              File "D:\python\lib\site-packages\torch\nn\modules\module.py",**kwargs)
              File "D:\python\lib\site-packages\torch\nn\modules\conv.py",line 423,in forward
                return self._conv_forward(input,self.weight)
              File "D:\python\lib\site-packages\torch\nn\modules\conv.py",line 419,in _conv_forward
                return F.conv2d(input,weight,self.bias,self.stride,RuntimeError: expected scalar type Double but found Float

我认为 x 有问题，我用 print(x.dtype) 打印它的类型：

torch.float64

这是双精度而不是浮点数。你知道怎么回事吗？感谢您的帮助！

解决方法

该错误实际上是指在调用矩阵乘法时默认情况下 float32 中的 conv 层的权重。由于您的输入是 double(pytorch 中的 float64) 而 conv 中的权重是 float
所以你的情况的解决方案是：

def train_net():
    dataset = ld.load()
    data_iter = Data.DataLoader(dataset,batch_size=168,shuffle=True)
    net = model.VGG_19()
    summary(net,(3,224,224),device="cpu")
    loss_func = nn.CrossEntropyLoss()
    optimizer = optim.SGD(net.parameters(),lr=0.001,momentum=0.9,dampening=0.1)
    scheduler = torch.optim.lr_scheduler.StepLR(optimizer,step_size=50,gamma=0.1)
    for epoch in range(5):
        print("epoch:",epoch + 1)
        train_loss = 0
        for i,data in enumerate(data_iter,0):
            x,y = data
            x = x.float()    # HERE IS THE CHANGE
            print(x.dtype)
            optimizer.zero_grad()
            out = net(x)
            loss = loss_func(out,y)
            loss.backward()
            optimizer.step()
            train_loss += loss.item()
            if i % 100 == 99:
                print("loss:",train_loss / 100)
                train_loss = 0.0
    print("finish train")

这肯定会奏效