machine-learning - Pytorch 中的预期双张量(获得浮点张量)

Question

我想在Pytorch中创建nn.Module 。我使用以下代码来解决文本相关问题(事实上，我使用 Glove 300d 预训练嵌入和句子中单词的加权平均来进行分类)。

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv1d(300, 128, kernel_size=5)
        self.conv2 = nn.Conv1d(128, 64, kernel_size=2)
        self.conv2_drop = nn.Dropout()
        self.fc1 = nn.Linear(64, 20)
        self.fc2 = nn.Linear(20, 2)

    def forward(self, x):
        x = F.relu(F.avg_pool1d(self.conv1(x), 2))
        x = F.relu(F.avg_pool1d(self.conv2_drop(self.conv2(x)), 2))
        x = x.view(-1, 1)
        x = F.relu(self.fc1(x))
        x = F.dropout(x, training=self.training)
        return self.fc2(x)

但它给了我以下错误:

Traceback (most recent call last):
    x = F.relu(F.avg_pool1d(self.conv1(x), 2))
  File "/Library/Frameworks/Python.framework/Versions/3.5/lib/python3.5/site-packages/torch/nn/modules/module.py", line 224, in __call__
    result = self.forward(*input, **kwargs)
  File "/Library/Frameworks/Python.framework/Versions/3.5/lib/python3.5/site-packages/torch/nn/modules/conv.py", line 154, in forward
    self.padding, self.dilation, self.groups)
  File "/Library/Frameworks/Python.framework/Versions/3.5/lib/python3.5/site-packages/torch/nn/functional.py", line 83, in conv1d
    return f(input, weight, bias)
RuntimeError: expected Double tensor (got Float tensor)

我对 Conv1d 相当陌生，大多数教程都使用 Conv1d 来解决图像问题。有人可以告诉我问题出在哪里吗？

我还在forward方法中添加了model.double()，但又出现了另一个错误:

RuntimeError: Given input size: (300 x 1 x 1). Calculated output size: (128 x 1 x -3). Output size is too small

Answer 1

错误1

RuntimeError: expected Double tensor (got Float tensor)

当您将双张量传递给第一个 conv1d 函数时，就会发生这种情况。 Conv1d 仅适用于浮点张量。要么这样做，

1. conv1.double() 或
2. model.double()。

这就是你所做的，而且是正确的。

错误2

RuntimeError: Given input size: (300 x 1 x 1). Calculated output size: (128 x 1 x -3). Output size is too small

这是因为您传递的输入窗口大小为 5 的卷积无效。您必须向 Conv1d 添加填充才能使其正常工作，如下所示:

self.conv1 = nn.Conv1d(300, 128, kernel_size=5, padding=2)

如果您不想添加填充，则给定 (batch_size, in_channels, inp_size) 作为输入张量的大小，您必须确保您的 inp_size 大于 5。

所有修复合并

确保您的尺寸对于网络的其余部分来说是正确的。就像这样:

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv1d(300, 128, kernel_size=5, padding=2)
        self.conv2 = nn.Conv1d(128, 64, kernel_size=2, padding=1)
        self.conv2_drop = nn.Dropout()
        self.fc1 = nn.Linear(64, 20)
        self.fc2 = nn.Linear(20, 2)

    def forward(self, x):
        x = F.relu(F.avg_pool1d(self.conv1(x), 2, padding=1))
        x = F.relu(F.avg_pool1d(self.conv2_drop(self.conv2(x)), 2))
        x = x.view(1, -1) # bonus fix, Linear needs (batch_size, in_features) and not (in_features, batch_size) as input.
        x = F.relu(self.fc1(x))
        x = F.dropout(x, training=self.training)
        return self.fc2(x)

if __name__ == '__main__':

    t = Variable(torch.randn((1, 300, 1))).double() # t is a double tensor
    model = Net()
    model.double() # this will make sure that conv1d will process double tensor
    out = model(t)