I found an awesome kernel on Hypercolumns for medical image segmentation on kaggle. Author of the kernel used fastai DynamicUnet as the model and added hypercolumns in it. Although I somewhat understand Hypercolumns in theory, trying to understand what he did in the kernel has left me perplexed.
Here is the kernel.
If someone could provide some pointers on what he did with the DynamicUnet_Hcolumns class, it would be a huge help.
I know this is a huge favor to ask. I think understanding how to use this powerful segmentation technique will be helpful further down the road.
P.S. I tried using his code on Camvid and failed miserably.