如何在 Pytorch 中实现 dropout，以及在哪里应用它

Question

我很不确定这是否正确。我真的很难找到很多关于如何参数化神经网络的好例子。

你如何看待这两个班级的辍学方式。首先，我正在编写原始类：

class NeuralNet(nn.Module):
  def __init__(self, input_size, hidden_size, num_classes, p = dropout):
      super(NeuralNet, self).__init__()
      self.fc1 = nn.Linear(input_size, hidden_size)
      self.fc2 = nn.Linear(hidden_size, hidden_size)
      self.fc3 = nn.Linear(hidden_size, num_classes)

  def forward(self, x):
      out = F.relu(self.fc1(x))
      out = F.relu(self.fc2(out))
      out = self.fc3(out)
      return out

然后在这里，我发现了两种不同的写东西的方式，我不知道如何区分。第一个使用：

self.drop_layer = nn.Dropout(p=p) 

而第二个：

self.dropout = nn.Dropout(p) 

这是我的结果：

class NeuralNet(nn.Module):
  def __init__(self, input_size, hidden_size, num_classes, p = dropout):
      super(NeuralNet, self).__init__()
      self.fc1 = nn.Linear(input_size, hidden_size)
      self.fc2 = nn.Linear(hidden_size, hidden_size)
      self.fc3 = nn.Linear(hidden_size, num_classes)
      self.drop_layer = nn.Dropout(p=p)

  def forward(self, x):
      out = F.relu(self.fc1(x))
      out = F.relu(self.fc2(out))
      out = self.fc3(out)
      return out


 class NeuralNet(nn.Module):
  def __init__(self, input_size, hidden_size, num_classes, p = dropout):
      super(NeuralNet, self).__init__()
      self.fc1 = nn.Linear(input_size, hidden_size)
      self.fc2 = nn.Linear(hidden_size, hidden_size)
      self.fc3 = nn.Linear(hidden_size, num_classes)
      self.dropout = nn.Dropout(p) 

  def forward(self, x):
      out = F.relu(self.fc1(x))
      out = F.relu(self.fc2(out))
      out = self.fc3(out)
      return out

这是否可行，如果不是如何改进它，它是否给了我我期望的结果，这意味着创建一个神经网络，我可以在其中删除一些神经元。重要的细节，我只想dropout第二层神经网络，其余不涉及！

Answer 1

您提供的两个示例完全相同。self.drop_layer = nn.Dropout(p=p)并且self.dropout = nn.Dropout(p)只是因为作者将层分配给不同的变量名称而有所不同。dropout 层通常在.__init__()方法中定义，并在 中调用.forward()。像这样：

 class NeuralNet(nn.Module):
  def __init__(self, input_size, hidden_size, num_classes, p = dropout):
      super(NeuralNet, self).__init__()
      self.fc1 = nn.Linear(input_size, hidden_size)
      self.fc2 = nn.Linear(hidden_size, hidden_size)
      self.fc3 = nn.Linear(hidden_size, num_classes)
      self.dropout = nn.Dropout(p) 

  def forward(self, x):
      out = F.relu(self.fc1(x))
      out = F.relu(self.fc2(out))
      out = self.dropout(self.fc3(out))
      return out

您可以进行以下测试：

import torch
import torch.nn  as nn

m = nn.Dropout(p=0.5)
input = torch.randn(20, 16)
print(torch.sum(torch.nonzero(input)))
print(torch.sum(torch.nonzero(m(input))))

tensor(5440) # sum of nonzero values
tensor(2656) # sum on nonzero values after dropout

让我们想象一下：

import torch
import torch.nn as nn
input = torch.randn(5, 5)
print(input)

tensor([[ 1.1404,  0.2102, -0.1237,  0.4240,  0.0174],
        [-2.0872,  1.2790,  0.7804, -0.0962, -0.9730],
        [ 0.4788, -1.3408,  0.0483,  2.4125, -1.2463],
        [ 1.5761,  0.3592,  0.2302,  1.3980,  0.0154],
        [-0.4308,  0.2484,  0.8584,  0.1689, -1.3607]])

现在，让我们应用 dropout：

m = nn.Dropout(p=0.5)
output = m(input)
print(output)

tensor([[ 0.0000,  0.0000, -0.0000,  0.8481,  0.0000],
        [-0.0000,  0.0000,  1.5608, -0.0000, -1.9459],
        [ 0.0000, -0.0000,  0.0000,  0.0000, -0.0000],
        [ 0.0000,  0.7184,  0.4604,  2.7959,  0.0308],
        [-0.0000,  0.0000,  0.0000,  0.0000, -0.0000]])

大约一半的神经元已经变为零，因为我们有可能p=0.5将神经元设置为零！