How can I fine tune last n layers not just a head?

[sh1ng]

Hi,

It would be nice to have not only be able to tune head layers, but a few last conv layers also.
Currently it's only postable to do

net = geffnet.create_model('efficientnet_b3')

....

def forward(self, x):
      with torch.no_grad():
          x = net.features(x)
       return head(x)

[rwightman]

@sh1ng right now you have to iterate over the model modules manually (model specific) to freeze BN, set requires_grad = False, or split params into different optimizer groups. I have a todo to make a generic interface to do this based on the feature extraction stages, but it's behind a bunch of other tasks.

In this aspect, it's no different from torchvision models.