Making an Age detector, need some helps

I am right now trying to create an age detector by referencing human faces. I decided to use pretrained model from resnet50. However, when I refer to other age detectors found from github, I found out they did import a pretrained model, but do not freeze the CNN layers.

I thought that since CNN layers are kind of feature extractors, it is reasonable to freeze those layers. If so, what are the reasons for not freezing those CNN layers??

I really need you guys’ help! Trying to learn everything from scratch.

Sometimes one needs to update the feature extractors too. That is the feature extractor for imagenet might not be optimal for age detection. If you go through fastai course, you will see that they first train the classifier by freezing the feature extractor in first stage. Then in the second stage they unfeeze the whole network and train all the layers.