How can I get the middle layer’s output of the trained model?

1

I have a train model,the architecture of the model as shown in the following:

def forward(self, x):
      x = F.relu(F.max_pool2d(self.conv1(x), 2))
      x = F.relu(F.max_pool2d(self.conv2_drop(self.conv2(x)), 2))
      x = F.relu(F.max_pool2d(self.conv3(x), 2))
      x = F.relu(F.max_pool2d(self.conv4_drop(self.conv4(x)), 2))

      x = x.view(-1, 160)

      return x

How can I get the middle layer’s output of in the test processing?

Can anyone give me some suggestions?
Thank you so much.

2

You don’t have to define all of your layers to be x. Try something like the following:

def forward(self, x):
      x1 = F.relu(F.max_pool2d(self.conv1(x), 2))
      x2 = F.relu(F.max_pool2d(self.conv2_drop(self.conv2(x1)), 2))
      x3 = F.relu(F.max_pool2d(self.conv3(x2), 2))
      x4 = F.relu(F.max_pool2d(self.conv4_drop(self.conv4(x3)), 2))

      x5 = x4.view(-1, 160)

      return x5

This way you can get at the output of the middle layers (x1 ... x4)

1 Like
3

This is very helpful,thank you.

5

variation on richard’s approach: simply return the xns you want, like:

return x4, x3, x2, x1

There is no rule/law that says that network modules need to return one and only one tensor.

(You can also return a list or a dictionary too by the way; I mean, you can return anything you like in fact; this is true for input parameters too by the way: you can have as many or as few input parameters as you want, and of any type you find works well for you)