Loading a pretrained model without freezing layers? Is it wrong to do?

vision
1

Hi guys, I’m using a resnet on a dataset with pretrained = True.

When I freeze the last layer, my model does not learn and plateau’s at 74% acc.

When I simply load the model without freezing any layer I get 81% acc.

Is it a wrong method to use?

2

Usually you would freeze the first layers and finetune the last one.
What’s your use case to only freeze the last layer?

1 Like
3

Hey @ptrblck, sorry for the late reply, you’re right.

I’m simply trying to use
model = models.resnet18(pretrained=True)

To classify images. When I freeze the layers my acc starts at 69%, when I don’t it starts at 74%.