Add Sequential model to Sequential

1

Hi I am new to Pyttorch and may be the question be simple,
I want to create Sequential model that have multiple Sequential segment or slice so that is my code

class Net(nn.Module):
      def __init__(self):
            super(Net, self).__init__()
             #input 3 32 32
            self.slic1=nn.Sequential()
            self.slic1.add_module('conv1', nn.Conv2d(3, 6, 7,padding=3))
        
            self.layers = nn.Sequential()
            self.layers.add_module('slic1',self.slic1)
            self.layers.add_module('Act1',torch.nn.ReLU())
           self.layers.add_module('pool1',nn.MaxPool2d(2, 2))
     
           self.layers.add_module('conv2', nn.Conv2d(6, 16 ,5,padding=2))
           self.layers.add_module('Act2',torch.nn.ReLU())
           self.layers.add_module('pool2',nn.MaxPool2d(2, 2))

           self.layers.add_module('flat',nn.Flatten())
           self.layers.add_module('lin1',nn.Linear(16 * 8 * 8, 500))
           self.layers.add_module('Act4',torch.nn.ReLU())
          self.layers.add_module('lin3',nn.Linear(500, 10))
  
     def forward(self, x):
          x_out = self.layers(x)
         return x_out

But conv layer at self.slic1 TWO time added to main model !!! when execute
net = Net()
weightlist=net.state_dict()
weightlist HAVE TWO time conv1 layer (layer defined in self.slic1) parameter,
what Is wrong ? and how I can create Sequential model with multiple segment

2

Hi,

This is because you have slic1 both in self. and self.layers.
If you only want it in self.layers, you should not create it as an attribute of self:

slic1=nn.Sequential()
slic1.add_module('conv1', nn.Conv2d(3, 6, 7,padding=3))
        
self.layers = nn.Sequential()
self.layers.add_module('slic1',slic1)
1 Like
3

Thank you :grinning:,
yes my problem solved but I want to access model slice from outside and attach to them
specific loss function in addition to model loss function, slice is not accessible from outside in this way

4

You can still access slice from model.layers.slic1.
If you want to save a nn.Module into another one, without it being detected as a child module, you can “hide” it by putting it into a list/dict: self.slic1 = [slic1,] and access it as self.slic1[0].