Use of nn.ReflectionPad2d(1)

I was looking for simple example of Siamese Networks and I found this article at hackernoon.

input batch first go through nn.ReflectionPad2d(1) as shown below:

class SiameseNetwork(nn.Module):
    def __init__(self):
        super(SiameseNetwork, self).__init__()
        self.cnn1 = nn.Sequential(
            nn.ReflectionPad2d(1),
            nn.Conv2d(1, 4, kernel_size=3),
            nn.ReLU(inplace=True),
            nn.BatchNorm2d(4),
            nn.Dropout2d(p=.2),
            
            nn.ReflectionPad2d(1),
            nn.Conv2d(4, 8, kernel_size=3),
            nn.ReLU(inplace=True),
            nn.BatchNorm2d(8),
            nn.Dropout2d(p=.2),

            nn.ReflectionPad2d(1),
            nn.Conv2d(8, 8, kernel_size=3),
            nn.ReLU(inplace=True),
            nn.BatchNorm2d(8),
            nn.Dropout2d(p=.2),
        )

        self.fc1 = nn.Sequential(
            nn.Linear(8*100*100, 500),
            nn.ReLU(inplace=True),

            nn.Linear(500, 500),
            nn.ReLU(inplace=True),

            nn.Linear(500, 5)
        )

    def forward_once(self, x):
        output = self.cnn1(x)
        output = output.view(output.size()[0], -1)
        output = self.fc1(output)
        return output

    def forward(self, input1, input2):
        output1 = self.forward_once(input1)
        output2 = self.forward_once(input2)
        return output1, output2
  1. why do siamese network (or the author’s approach) need padding before feeding into convolutions ?
  2. what is the point of using ReflectionPad instead of zero padding ?
  1. I guess the ReflectionPad2d layers were added as nn.Conv2d supported zero padding only in the past (in new PyTorch versions you can specify the padding_mode).
  2. I don’t know if the author has explained this architecture in a research paper, but would guess that this padding type worked better than zero padding based on their experiments.
1 Like