feat: add dice loss in linknet #816
Conversation
charlesmindee
added
type: enhancement
Codecov Report
@@ Coverage Diff @@
## main #816 +/- ##
==========================================
+ Coverage 96.01% 96.02% +0.01%
==========================================
Files 131 131
Lines 4941 4935 -6
==========================================
- Hits 4744 4739 -5
+ Misses 197 196 -1
Flags with carried forward coverage won't be shown. Click here to find out more.
Continue to review full report at Codecov.
|
| inter = (prob_map[seg_mask] * seg_target[seg_mask]).sum() | ||
| cardinality = (prob_map[seg_mask] + seg_target[seg_mask]).sum() | ||
| dice_loss = 1 - 2 * inter / (cardinality + 1e-8) |
There was a problem hiding this comment.
we'll have to update this in another PR because it highly advantages classes that occupy the biggest area
There was a problem hiding this comment.
Yes, I will do it in the next PR !
There was a problem hiding this comment.
To curb this, may be FL + Dice Loss could be a better option? Because, even if the area was same for the classes, since total area of combined bboxes / total area of image could be really less, it may tend to learn more negatives.
There was a problem hiding this comment.
Yup, I meant that the baseline option would be to average the loss over classes (not mixing the contribution of each class to the loss before the last step). But your suggestion @SiddhantBahuguna is worth exploring as a follow-up and could improve results 👍
There was a problem hiding this comment.
Thanks for your suggestions, I will add focal loss in the next PR !
This PR adds dice loss in the linknet loss (which is now bce_loss + dice_loss) and remove the boosted edges part of the loss since it hasn't led to successful trainings. Any feedback is welcome!