Inference and predictions¶

Run inference on a validation split and save results for metrics or visualization.

CLI¶

# Checkpoint/config auto-resolve from experiment-dir when omitted
odet preds --experiment-dir runs/oriented_rcnn/<timestamp> --data-split val

Or from the repo root after training:

make preds
make metrics

Inference-only (skip mAP / analysis):

odet preds --experiment-dir runs/oriented_rcnn/<timestamp> --data-split val --no-diagnostics

Behavior¶

Thresholds, NMS, and sliding-window overlap come from production.* in the experiment config.json
Images larger than the model canvas use sliding-window tiling in save_predictions.py, then merge detections in image space

Output artifacts¶

Default directory: predictions/<YYYYMMDD_HHMMSS>/ at the repository root (or --output-dir).

File	When	Contents
`predictions.json`	Always (inference)	Per-image detections (rboxes, scores, labels)
`analysis_iou0.50.json`	With diagnostics (default)	PR/F1 threshold sweep, per-class AP, confusion matrix, best-threshold block
`model_analysis_<timestamp>.md`	With diagnostics	Human-readable report: per-class gts/dets/recall/AP, optional per-class best thresholds
`tile_metrics.csv`	With `--save-tile-metrics-csv`	Per-tile precision/recall/F1 for hard-tile oversampling
`visualizations/`	With `--save-visualizations`	Overlay images

Per-class best-threshold tables are computed by default. Disable with --no-per-class-threshold-analysis (--per-class-threshold-analysis is kept for backward compatibility).

Recompute metrics only¶

odet preds --metrics-from-json predictions/<timestamp> --config runs/.../config.json

Re-runs mAP and writes fresh analysis_iou*.json / model_analysis_*.md from an existing predictions.json without re-inference.

See the tools reference on GitHub and Getting Started: Tools.