How to add an additional layer on top of a pretrained model?

1

How can I append an additional layer on top of an existing pretrained architecture ? I saw this snippet below from Using freezed pretrained resnet18 as a feature extractor for cifar10 but it modifies the last layer instead of appending a new one to it.

model = torchvision.models.vgg19(pretrained=True)
for param in model.parameters():
    param.requires_grad = False
    # Replace the last fully-connected layer
    # Parameters of newly constructed modules have requires_grad=True by default
model.fc = nn.Linear(512, 8) # assuming that the fc7 layer has 512 neurons, otherwise change it 
model.cuda()

In a later part of the code, the model was called as follows when processing the input data

output = model(input)

2 Likes
2

A very simple way is

model = nn.Sequential(models.vgg19(True), YourModule())
# Use it as usual
out = model(input)
13 Likes
3

Hi,
Also is it possible to add one conv layer to vgg19 in this way?
Thanks!

4

hi. What is “YourModule()”? Could you pls explain to me?

5

YourModule() refers to any function that returns a nn.Module. So that’s just a placeholder to say that you can insert whatever model you want in there.

1 Like
6

In this snippet how do we make sure that the output dimension models.vgg19 is in sync with the input dimension of YourModule()?

1 Like
7

I have the same problem.

8

:pray: :pray: :pray: Thanks a lot RicCu. I was stuck in trying to modify the original end-classifier outputting 600 features in a hybrid classical-quantum transfer learning model. I needed a binary output. I used : model_hybrid = nn.Sequential(model, DressedQuantumNet()), and it worked!