Freezing feature detection layers

[Swearinbag]

I've implemented a way to freeze layers of the network but I wonder what interval of layers decide the feature extraction of the input, so that I can freeze these and see if the model performance increases depending on my custom dataset.