Is there a specific reason why layer freezing during pretraining is not allowed. I was thinking of doing pretraining with ImageNet and taking that pre-trained network and pretraining the last block with my own unlabelled dataset which is <18K images.