fix: Fixed LinkNet target & loss computation

[fg-mindee]

This PR introduces the following modifications:

• fixed edge_mask building
• fixed seg_target for PyTorch (was built as channels last)
• removes focal_loss option in LinkNet loss computation

Any feedback is welcome!

[codecov]

Codecov Report

Merging #670 (11167ed) into main (1dad108) will increase coverage by 0.08%.
The diff coverage is 97.14%.

@@            Coverage Diff             @@
##             main     #670      +/-   ##
==========================================
+ Coverage   96.17%   96.26%   +0.08%     
==========================================
  Files         117      117              
  Lines        4472     4467       -5     
==========================================
- Hits         4301     4300       -1     
+ Misses        171      167       -4     

Flag	Coverage Δ
unittests	96.26% <97.14%> (+0.08%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
doctr/models/detection/linknet/base.py	89.02% <94.44%> (+1.02%)	⬆️
...dels/detection/differentiable_binarization/base.py	91.30% <100.00%> (-0.63%)	⬇️
doctr/models/detection/linknet/pytorch.py	98.03% <100.00%> (+1.77%)	⬆️
doctr/models/detection/linknet/tensorflow.py	100.00% <100.00%> (+3.48%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 1dad108...11167ed. Read the comment docs.

[charlesmindee]

Thanks, but why do you remove the focal loss ?

[fg-mindee]

Thanks, but why do you remove the focal loss ?

It was introducing an asymetry in model API, and we added it because we were having bad performances with LinkNet initially. So I figured, it's better to start again simple, make it work and then if we decide that all segmentation models need it, we'll reintroduce it 👍

[charlesmindee]

Thanks!