WIP: Protocol-defined neural network inference

[gwennlbh]

I've started something, but this probably needs a discussion with @edgaremy and @axelcarlier before settling on the data we ask to protocol definitions

The end-goal here is possibly to get rid of built-in metadata and protocols all together (or at least making the built-in "default" protocol a regular protocol defined in a .yaml file)

this potentially requires a lot of de-harcoding in a codebase but i think it's worth it, especially since i suppose the neural net (espeically on the classification part) will improve over time and potentially grow with more classes

This PR targets the sidepanelredesign branch that contains taxonomy tree changes (see #95), since classification logic, species metadata definition and taxonomic inference are deeply intertwined.

TODO:

• Take into account that output detection shape can have cx/ cy instead of sx/sy. Currently we act as if sx=cx and sy=cy (or the reverse, can't remember)
• Allow specifying ex/ey instead of w/h (end x and end y)
• Figure out what our model really has as data instead of w and h (theres a weird $w\leftarrow |x-w|$ and similarily for h, w & h are probably ex and ey in our case instead)
• Take into account output.normalized et input.normalized dans tout les cas
• Take into account input.disposition in all cases

[github-actions]

PR Preview Action v1.6.0
🚀 View preview at https://cigaleapp.github.io/cigale/_pullrequests/pr-99/
Built to branch gh-pages at 2025-03-22 02:41 UTC. Preview will be ready when the GitHub Pages deployment is complete.