Transfer Learning in plain PyTorch using Fastai Insights

Hey guys. Recently I was trying to use the knowledge from the fastai course in my own experiments since I prefer working in Vanilla PyTorch and implementing stuff myself.

The above post resulted from those experiments. How to use insights from fastai’s transfer Learning process (which gives SOTA almost all of the time) in plain PyTorch.

Discriminative Lr, don’t freeze batchnorm, custom head etc.