Torchvision vgg16 pretrained model layer naming

1

Hello,

I am curious about the layer naming (key values of state_dict) of the vgg16 pretrained model from torchvision.models module, e.g.: ‘features.0.weight’, ‘features.0.bias’, ‘features.2.weight’, ‘features.2.bias’, etc. The number increases by 2 with each new convolutional or fc layer, and increases by 3 each time encountering a max pooling layer.

Do you have an idea of the underlying logic?

Thank you very much!

Guangye

2

The numbered indices in these modules names are created by the nn.Sequential module.
Layers without any parameters will still get the index, but won’t be shown in the state_dict.
Here is a small example:

model = nn.Sequential(
    nn.Linear(1, 1),
    nn.ReLU(),
    nn.Linear(1, 1),
    nn.ReLU(),
    nn.Sigmoid(),
    nn.Linear(1, 1))
print(model.state_dict().keys())
> odict_keys(['0.weight', '0.bias', '2.weight', '2.bias', '5.weight', '5.bias'])

As you can see, only the layer indices 0, 2, and 5 are shown, as these layers contain parameters.