Sorry for the late response Sanchit Tanwar .

Part 1 of your Qs:-

We are using precompute=true initially for sure but later on we are unfreezing the whole architecture and updating the parameters .

Part 2 of your Qs:-

Resnet has some clear advantage over VGG , hence the accuracy achieved by resnet will be better than VGG.

Part 3 of your Qs:-

If you go under the hood , i.e explore the code for each function using (??function_name) , you will find that its not simply applying resnet architecture over the dataset but Jeremy has come up with lots of interesting techniques that are being used to get that State of the art results. You might not have applied those in your Keras implementation.

To understand those codes on top of which fastai exists , I recommend you to learn basics of Python OOPs programing and Pytorch .

I bet you the outcome will be more satisfying .

P.S- Great work there doing the keras implementation, keep experimenting . That curiosity will get you most of your answers.