The model.parameter() feed into optimizer when backpropagation

[lisyuan]

I want to frozen the weight of pretrained vgg19, and just update the last fc layer when backpropagation
The model is shown in following block. I have already set with torch.no_grad(): when I input the images into vgg19.

class ImgEncoder(nn.Module):

    def __init__(self, embed_dim):

        super(ImgEncoder, self).__init__()
        self.model = models.vgg19(pretrained=True)
        in_features = self.model.classifier[-1].in_features
        self.model.classifier = nn.Sequential(*list(self.model.classifier.children())[:-1]) # remove vgg19 last layer
        self.fc = nn.Linear(in_features, embed_dim)

    def forward(self, image):

        with torch.no_grad():
            img_feature = self.model(image) # (batch, channel, height, width)
        img_feature = self.fc(img_feature)

        l2_norm = F.normalize(img_feature, p=2, dim=1).detach()
        return l2_norm

model = ImgEncoder(embed_dim=300)

the question is when I using the optimizer, how should I feed the parameter

optimizer = optim.Adam(model.parameters(), lr=0.001)

or

optimizer = optim.Adam(model.fc.parameters(), lr=0.001)

This question is similar to the tutorial in official website like following:

Why we can’t simply put the model.parameters() but model.fc.parameters() instead ?

[Nikronic (Nikan Doosti)]

Hi,

The answer is yes, you can and there won’t be any differences in result.
Please see this thread for more details:

Passing to the optimizers frozen parameters

When I want to do transfer learning, I set the require_grad = False and only pass the other parameters to the optimizer. But what happens if I only do one of these two steps? What happens if I set several modules’ reguire_grad = False but then I pass all net.parameters() to the optimize. What happens if I keep the require_grad = True but only pass a subset of the parameters to the optimizer (the subset I want to fine tune, for example).

Bests

[lisyuan]

thanks for your sharing!