python - 如何修复 : RuntimeError: size mismatch in pyTorch

Question

我是 pyTorch 的新手，遇到以下大小不匹配错误:

RuntimeError: size mismatch, m1: [7 x 2092500], m2: [180 x 120] at ..\aten\src\TH/generic/THTensorMath.cpp:961

型号:

class Net(nn.Module):
def __init__(self):
    super(Net, self).__init__()
    self.conv1 = nn.Conv2d(3, 200, 5)
    self.pool = nn.MaxPool2d(2, 2)
    self.conv2 = nn.Conv2d(200, 180, 5)
    self.fc1 = nn.Linear(180, 120)
    self.fc2 = nn.Linear(120, 84)
    self.fc3 = nn.Linear(84,5)     

def forward(self, x):
    x = self.pool(F.relu(self.conv1(x)))
    x = self.pool(F.relu(self.conv2(x)))
    x = x.view(x.shape[0], -1)
    x = F.relu(self.fc1(x))
    x = F.relu(self.fc2(x))
    x = self.fc3(x)
    return x

我是如何尝试将 x = x.view(x.shape[0], -1) 更改为 x = x.view(x.size(0), -1 ) 但这也不起作用。图像尺寸为 512x384。并使用了以下转换:

def load_dataset():
data_path = './dataset/training'

transform = transforms.Compose(
               [transforms.Resize((512,384)),
                transforms.ToTensor(),
                transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])


train_dataset = torchvision.datasets.ImageFolder(root=data_path,transform=transform)
train_loader = torch.utils.data.DataLoader(train_dataset,batch_size=7,num_workers=0,shuffle=True)

return train_loader

Answer 1

问题是最后一个最大池化层的输出维度与第一个全连接层的输入维度不匹配。这是输入形状 (3, 512, 384) 的最后一个最大池层之前的网络结构:

----------------------------------------------------------------
        Layer (type)               Output Shape         Param #
================================================================
            Conv2d-1        [-1, 200, 508, 380]          15,200
         MaxPool2d-2        [-1, 200, 254, 190]               0
            Conv2d-3        [-1, 180, 250, 186]         900,180
         MaxPool2d-4         [-1, 180, 125, 93]               0
================================================================

表的最后一行表示 MaxPool2d-4 输出 180 个 channel (滤波器输出)，宽度为 125，高度为 93。因此，您需要第一个完全连接的层具有 180 * 125 * 93 = 2092500 输入大小。这是很多，所以我建议你改进你的架构。在任何情况下，如果您将第一个全连接层的输入大小更改为 2092500，它就会起作用:

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(3, 200, 5)
        self.pool = nn.MaxPool2d(2, 2)
        self.conv2 = nn.Conv2d(200, 180, 5)
        #self.fc1 = nn.Linear(180, 120)
        self.fc1 = nn.Linear(2092500, 120)
        self.fc2 = nn.Linear(120, 84)
        self.fc3 = nn.Linear(84,5)

    def forward(self, x):
        x = self.pool(F.relu(self.conv1(x)))
        x = self.pool(F.relu(self.conv2(x)))
        x = x.view(x.shape[0], -1)
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        return x

给出以下架构:

----------------------------------------------------------------
        Layer (type)               Output Shape         Param #
================================================================
            Conv2d-1        [-1, 200, 508, 380]          15,200
         MaxPool2d-2        [-1, 200, 254, 190]               0
            Conv2d-3        [-1, 180, 250, 186]         900,180
         MaxPool2d-4         [-1, 180, 125, 93]               0
            Linear-5                  [-1, 120]     251,100,120
            Linear-6                   [-1, 84]          10,164
            Linear-7                    [-1, 5]             425
================================================================
Total params: 252,026,089
Trainable params: 252,026,089
Non-trainable params: 0

(您可以使用 torchsummary 包来生成这些表。)