Plot_top_losses for Unet

What is the best approach of deriving plot_top_losses for Unet segmentation? My goal is to find cases that are poorly classified or cases with lower confidence, so I can perform error analysis and possibly train further on a subset. If this is covered somewhere else I can certainly review that documentation!

cc: @sgugger

Segmentation Interpretation is already implemented with a top_losses https://github.com/fastai/fastai/blob/master/fastai/vision/interpret.py

If you wanted to plot, I’d look at the show code and perhaps visions plot top losses function for ideas (if the above isn’t enough for you!)

1 Like

Thanks! I didn’t realize this. Let me try it and see if addresses what I need.

Hi Akshay,

Any luck with plotting top losses for segmentation? Would you like to share some insights on how you achieved it.

Thanks in advance
Bilal