Pytorch 上的 1D CNN：mat1 和 mat2 形状不能相乘（10x3 和 10x2）

Question

我有一个时间序列，包含 500 个大小的样本和 2 种类型的标签，并且想要使用 pytorch 构建一个 1D CNN：

class Simple1DCNN(torch.nn.Module):
    def __init__(self):
        super(Simple1DCNN, self).__init__()
        self.layer1 = torch.nn.Conv1d(in_channels=50, 
                                      out_channels=20, 
                                      kernel_size=5, 
                                      stride=2)
        self.act1 = torch.nn.ReLU()
        self.layer2 = torch.nn.Conv1d(in_channels=20, 
                                      out_channels=10, 
                                      kernel_size=1)
        
        self.fc1 = nn.Linear(10* 1 * 1, 2)
    def forward(self, x):
        x = x.view(1, 50,-1)
        x = self.layer1(x)
        x = self.act1(x)
        x = self.layer2(x)
        x = self.fc1(x)
        
        return x

model = Simple1DCNN()
model(torch.tensor(np.random.uniform(-10, 10, 500)).float())

但收到此错误消息：

Traceback (most recent call last):
  File "so_pytorch.py", line 28, in <module>
    model(torch.tensor(np.random.uniform(-10, 10, 500)).float())
  File "/Users/lib/python3.8/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
  File "so_pytorch.py", line 23, in forward
    x = self.fc1(x)
  File "/Users/lib/python3.8/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/Users/lib/python3.8/site-packages/torch/nn/modules/linear.py", line 93, in forward
    return F.linear(input, self.weight, self.bias)
  File "/Users/lib/python3.8/site-packages/torch/nn/functional.py", line 1692, in linear
    output = input.matmul(weight.t())
RuntimeError: mat1 and mat2 shapes cannot be multiplied (10x3 and 10x2)

我究竟做错了什么？

Answer 1

x = self.layer2(x)该行的输出（也是下一行的输入x = self.fc1(x)）的形状是torch.Size([1, 10, 3])。

现在根据 的定义self.fc1，它期望输入的最后一个维度是 ，10 * 1 * 1而10您的输入3因此会出现错误。

我不知道你想做什么，但假设你想做的是；

1. 将整个500尺寸序列标记为两个标签之一，然后执行此操作。

# replace self.fc1 = nn.Linear(10* 1 * 1, 2) with
self.fc1 = nn.Linear(10 * 3, 2)

# replace x = self.fc1(x) with
x = x.view(1, -1)
x = self.fc1(x)

2. 将10每个时间步标记为两个标签之一，然后执行此操作。

# replace self.fc1 = nn.Linear(10* 1 * 1, 2) with
self.fc1 = nn.Linear(2, 2)

1的输出形状为 (batch size, 2)，2 的输出形状为 (batch size, 10, 2)。