Two solutions

Scheme 1 ： Check whether the network and data are GPU On

RuntimeError: Input type (torch.FloatTensor) And weight type (torch.cuda.FloatTensor) It should be the same , Or the input should be a MKLDNN tensor , And weight is a dense tensor .

Our weight is cuda type (GPU Train to get ), And input ( Data to be tested ) No cuda type . use GPU Training model , Not directly in CPU Upper use , Put it in GPU Medium forecast .

If the model parameters used are cuda On (gpu) Trained , When using it for testing , You need to put the data to be tested into GPU On , namely ：data.cuda().

Carefully check whether the neural network and related data are placed GPU On ！

Option two ： Model definitions are all placed in class initialization

Besides , According to the analysis of the boss , May also be , Because your model is defined , Not well defined , As a result, part of the model cannot be transferred to cuda On . The specific test is skipped .

in general , If the above method doesn't work , Then try putting all the network layers （ As long as there are parameters to be trained in the network layer ） in __init__() Function to define , Only in forward() Write the logic of the runtime , namely ：

class A(nn.Module):
    def __init__(self):
        super(A,self).__init__()
        self.conv = nn.Conv2d(in_channels=3, out_channels=8, kernel_size=3)
        self.relu = nn.ReLU(inplace=True)
        self.b_module = B()

    def forward(self,x):
        out = self.conv(x)
        out = self.relu(out)
        out = self.b_module(out)
        return out

class B(nn.Module):
    def __init__(self):
        super(B,self).__init__()
        self.conv = nn.Conv2d(in_channels=8, out_channels=16, kernel_size=3)
        self.relu = nn.ReLU(inplace=True)

    def forward(self, x):
        out = self.conv(x)
        out = self.relu(out)
        return out

Reference resources ：

• Pytorch Avoid the pit ：RuntimeError: Input type(torch.cuda.FloatTensor) and weight type(torch.FloatTensor) shoul