Using one trained model as a extractor in another model

vision
1

I am using a trained model as an extractor in another model and it seems like it is working ok, but I don’t know if I am doing it right or now. So this is my model and how it is declared.

class MyModel(nn.Module):
    def __init__(self):
        super(MyModel, self).__init__()
        image_modules = list(ae_model.children())[:-2] #get only the encoder layers
        self.modelA = nn.Sequential(*image_modules)
        # Shape of max pool = 4, 112, 112
        self.fc1 = nn.Linear(4*112*112, 500)
        self.fc2 = nn.Linear(500,500)
        self.out = nn.Linear(500, 10)
        
        self.drop = nn.Dropout(0.3)
        
    def forward(self, x):
        x = self.modelA(x)
        x = x.view(x.size(0),4*112*112)
        x = torch.relu(self.fc1(x))
        x = self.drop(x)
        x = torch.relu(self.fc2(x))
        x = self.drop(x)
        x = self.out(x)
        return x

Now I have done this for a Resnet50 model but it did not look like this. My extractor that I am using is this

# define the NN architecture
class ConvAutoencoder(nn.Module):
    def __init__(self):
        super(ConvAutoencoder, self).__init__()
        ## encoder layers ##
        self.conv1 = nn.Conv2d(3, 16, 3, padding=1)  
        self.conv2 = nn.Conv2d(16, 4, 3, padding=1)
        self.pool = nn.MaxPool2d(2, 2)
        
        ## decoder layers ##
        self.t_conv1 = nn.ConvTranspose2d(4, 16, 2, stride=2)
        self.t_conv2 = nn.ConvTranspose2d(16, 3, 2, stride=2)
    def forward(self, x):
        ## encode ##
        x = torch.relu(self.conv1(x))
        x = self.pool(x)
        x = torch.relu(self.conv2(x))
        x = self.pool(x)  # compressed representation
        ## decode ##
        x = torch.relu(self.t_conv1(x))
        x = torch.sigmoid(self.t_conv2(x))
        return x

I think this is working but I would like to get a second opinion on it.
Thank you for any help with this.

2

I assume that ae_model is an instance of ConvAutoencoder.
If that’s the case, note that self.modelA will only use the child modules of ConvAutoencoder and will not apply the functional API calls in its forward such as torch.relu and torch.sigmoid and will also use the max pooling layer once.

nn.Sequential is a simple container to call modules in a sequential way. You could e.g. add the activation operations as modules (nn.ReLU() and nn.Sigmoid()) to the sequential model.
I think the better way would be to create a new model with the desired submodules and just call it in MyModel.

3

I see I should explain. I want just the encoder part so I want to output of the maxpooling layer only. So would what I have run the image through the encoder part and output the maxpooling layer part. From printing the output I think that is what is happening.
Thank you again.