Analysefunktionen

[florinbarbisch]

Florin: vizualization metrics Yvo: classical ML metrics

• [x] Majority-voting @florinbarbisch
• [x] Confusion matrices @yvokeller
• [x] Other Metrics @yvokeller @florinbarbisch
• [x] Map mit false classification @florinbarbisch

Metrics we could implement (from AgroLuege):

• [x] Single scores (before field majority) per level:
  • [x] overall_accuracy_bevor_field_majoritylevel{level}
  • [x] kappa_bevor_field_majoritylevel{level}
  • [x] precision_bevor_field_majoritylevel{level}
  • [x] recall_bevor_field_majoritylevel{level}
  • [x] f1_bevor_field_majoritylevel{level}
  • [x] perclassacc_bevor_field_majoritylevel{level}
• [x] Single scores (after field majority) per level:
  • [x] overall_accuracylevel{level}
  • [x] kappalevel{level}
  • [x] precisionlevel{level}
  • [x] recalllevel{level}
  • [x] f1level{level}
  • [x] perclassacclevel{level}
• [x] Pixel Accuracylevel{level}
• [x] confusion matrix_bevor_field_majoritylevel{level}
• [x] confusion matrix_level_{level}": wandb.Image(plt)
• [x] Example Fields bevor field majority
• [x] Example Fields
• [ ] ~log_wandb = dict(zip(label_names + f"_level_bevor_field_majority_{level}", cl_acc))~
• [ ] ~log_wandb = dict(zip(label_names + f"_level_{level}", cl_acc))~

PLZ: don't use the name before field majority but rather name the metrics with field majority kappa_field_majority_level_3

TODO:

• [x] Probably it makes sense to log per_class_accuracies as a table (will lose history though). Currently we have already over 200 metrics being tracked...
• [x] Looks like W&B web app can't handle our big confusion matrices well. Might need to opt for matplotlib plot for the big ones.

[yvokeller]

@florinbarbisch Han Accuracy, Overall Accuracy, Per Class Accuracy, Precision, Recall, F1, Kappa und Confusion Matrix implementiert. Het paar refactors braucht zum's schön anebecho, aber hans jetzt sinnvoll mit Callbacks chöne löse (class LogConfusionMatrix(pl.Callback) und class LogMessisMetrics(pl.Callback))

Chasch gern uf dem aufbauend Majority Voting etc. iibaue :)

[florinbarbisch]

Alli metrics si implemented. Aber s geit glaub huere lang zum die berechne😅

[yvokeller]

Uff ja han denkt das chönnt evtl. no chli es Problem werde... Beobachted mer's mal wenn mer denn en längere run mached mit allne Date

[florinbarbisch]

• [x] original besser benennen
• [x] alle bilder sollten prefix train_images_ haben
• [x] confusion matrix => schrift weiss maken
• [x] false classifcation (grün/rot karte) => background false classification hellrot einfärben
• [x] WandB captions für Bilder