DoB24's picture
Add post-hoc analysis: confusion matrices, ROC, reliability, per-class F1, accuracy ranking
add299a verified
|
Raw
History Blame Contribute Delete
708 Bytes
# Analysis Artefacts
Generated post-hoc from the test-set prediction JSONs (`results/*_test_preds.json`)
and the ensemble report.
## Contents
- `confusion_matrices/cm_<model>.png` β€” per-class confusion matrix (counts overlaid on row-normalised heatmap)
- `roc_curves/roc_<model>.png` β€” one-vs-rest ROC for every class + macro-AUC in title
- `reliability_diagrams/reliability_<model>.png` β€” 10-bin reliability with ECE in title
- `per_class/per_class_f1.png` β€” grouped bar chart, all 9 backbones side-by-side
- `per_class/per_class_f1.json` β€” raw per-class F1 numbers
- `accuracy_ranking.png` β€” horizontal bar chart of test accuracy
- `ece_summary.json` β€” Expected Calibration Error per model