I want to modify the pre-trained model and train it.
For example, I use the resnet50 model in torchvision.
I cancel the last fc layer
Then I use a new transposed convolution layer and a new maxpool layer and a new classifier layer.
My question: are these three new layers’ parameters already initialized?
Or I need to initialize them with method like “Xavier” ?
I checked the values of these three layers, and they are not zero. I don’t know whether they are already initialized.
And why they are not equal to zero? Where do these values come from?