运行用顺序构建的Pytorch CNN时，我收到一个错误：“形状不能相乘”，但我检查了形状是否匹配

Question

我非常困惑，为什么我有一个形状错误后，验证了层的输出形状。有人能帮我找出我哪里错了吗？

根据我所包含的层的总结，在第6层和第7层之间似乎出现了错误，但是第6层的输出显示与第7层的输入相同的维度，应该注意的是，误差维度6272与第3/4层的输出相对应。

我收到了这个错误：

回溯(最近一次调用)：

  File "C:\Users\logan\Spyder_ProjectCode.py", line 215, in <module>
    training_loss[t] = train_loop(trainloader, model, loss_fn, opt)/len(trainloader)

  File "C:\Users\logan\Spyder_ProjectCode.py", line 175, in train_loop
    pred = model(X)

  File "C:\Users\logan\anaconda3\lib\site-packages\torch\nn\modules\module.py", line 1110, in _call_impl
    return forward_call(*input, **kwargs)

  File "C:\Users\logan\anaconda3\lib\site-packages\torch\nn\modules\container.py", line 141, in forward
    input = module(input)

  File "C:\Users\logan\anaconda3\lib\site-packages\torch\nn\modules\module.py", line 1110, in _call_impl
    return forward_call(*input, **kwargs)

  File "C:\Users\logan\anaconda3\lib\site-packages\torch\nn\modules\linear.py", line 103, in forward
    return F.linear(input, self.weight, self.bias)

RuntimeError: mat1 and mat2 shapes cannot be multiplied (8x6272 and 1152x512)

#图层摘要

        Conv2d-1           [-1, 64, 15, 15]           1,792
          ReLU-2           [-1, 64, 15, 15]               0
        Conv2d-3            [-1, 128, 7, 7]          73,856
          ReLU-4            [-1, 128, 7, 7]               0
     MaxPool2d-5            [-1, 128, 3, 3]               0
       Flatten-6                 [-1, 1152]               0
        Linear-7                  [-1, 512]         590,336
          ReLU-8                  [-1, 512]               0
        Linear-9                  [-1, 340]         174,420
         ReLU-10                  [-1, 340]               0
       Linear-11                   [-1, 47]          16,027

================================================================

这是我的代码：

model = nn.Sequential(
    Conv2d(3, 64, kernel_size=3, stride=2),
    ReLU(),
    Conv2d(64, 128, kernel_size=3, stride=2),
    ReLU(),
    MaxPool2d((2,2), stride=(2,2)),
    Flatten(),
    Linear(3*3*128, 512),
    ReLU(),
    Linear(512, 340),
    ReLU(),
    Linear(340, 47)
    )

loss_fn = nn.CrossEntropyLoss()
learning_rate = 0.1
epochs = 15
momen = 0.8
model = model.to(device)       #choose one or the other
opt = optim.SGD(model.parameters(), lr=learning_rate, momentum=momen)

def train_loop(dataloader, model, loss_fn, optimizer):
    size = len(dataloader.dataset)
    training_loss = 0
    model.train()
    for batch, (X, y) in enumerate(dataloader):
        X, y = X.to(device), y.to(device)  
        pred = model(X)
        loss = loss_fn(pred, y)

        opt.zero_grad()
        loss.backward()
        opt.step()

        training_loss += loss.item()
    return training_loss

training_loss = np.zeros(epochs)
for t in range(epochs):
    print(f"Epoch {t+1}\\n-------------------------------")
    training_loss\[t\] = train_loop(trainloader, model, loss_fn, opt)/len(trainloader)
print("Done!")

Answer 1

这个错误是输入的数学错误。在运行火炬总结时，当训练期间的实际输入为(3,64,64)时，我给出的输入为(3,32,32)。因此，当我检查打印输出中各层之间的输出形状时，为什么没有检测到错误。我增加了一个额外的conv2d层，使输出形状达到所需的123,3,3，3进入完全连接的层。