# Varroa: three YOLOv8-P2 ASRM variants

Standalone package for training the three current ASRM architectures on the Varroa
dataset. It is isolated from LEVIR-Ship and from the original Varroa project.

Cases:

- `p2_asrm_fusion`
- `input_asrm_auxiliary`
- `input_asrm_guided_p2`

The notebook follows `Varroa/YOLO_custom/HF_YOLO_version_workflow.ipynb`:
`gt_one`, infected images only, class token 3 mapped to class 1, fixed conversion
split seed 42, pretrained `yolov8n.pt`, image size 640, batch 16, 100 epochs,
patience 20, workers 4 and training seeds 42/43.

The raw archive retains the original train/val/test layout. `prepare_dataset.py`
deduplicates the positive CSV records and creates a reproducible 70/15/15 Ultralytics
layout. All cases and training seeds use that same converted dataset.

Outputs:

- `test_metrics.json` per case/seed;
- `comparison_varroa_asrm.csv` with raw runs;
- `comparison_varroa_asrm_aggregate.csv` with mean/std over seeds.

Version 2 exposes an auxiliary attention map but has no auxiliary loss, so its
detection path is intentionally equivalent to the P2 baseline during this experiment.