Running through the same layer using a second variable

1

Can anyone help me what is the difference between these two codes? I mean what happens when we use the same layer for a second time by another variable. Any graphical model would be highly appreciated!

class Net(nn.Module):
def init(self):
super(Net, self).init()
self.conv1 = nn.Conv2d(1, 32, 3, 1)
self.conv2 = nn.Conv2d(32, 64, 3, 1)

def forward(self, x1,x2):
  x1 = self.conv1(x1)
  x2 = self.conv2(x2)
  out=x1+x2
  return output

class Net(nn.Module):
def init(self):
super(Net, self).init()
self.conv1 = nn.Conv2d(1, 32, 3, 1)

def forward(self, x1,x2):
  x1 = self.conv1(x1)
  x2 = self.conv1(x2)
  out=x1+x2
  return output
2

The difference would be in the parameters used. The first example has two convolutions each with their own parameters (weights) while the second shares parameters across the two inputs.

As is, I don’t think both versions can work at the same time because the the two convs in the first example have a different number of input and output channels implying that x1/x2 have different inputs channels. This means that in the second example self.conv(x2) isn’t possible. Summing x1 and x2 in the first example is also strange though it might be possible via broadcasting.

3

Thank you for your response. If we consider each generator as an encoder, the first one has two different encoders with different weights while the second one has two encoders with the shared weights. Am I correct? This is the concept of two encoders with shared weights that sometimes papers refer to it?