rosewt/petcare-derm-yolov8-cls-v0

PetCare Dermatology Triage — YOLOv8n-cls (v0, coarse baseline)

⚠️ DIRECTIONAL BASELINE ONLY — NOT a clinical or diagnostic model. This is an early v0 sanity-check baseline for the PetCare project. It was trained on a small, noisy, web-sourced dataset with coarse "lumped" labels and an 88-image validation split. Metrics below indicate the pipeline works end-to-end; they are not a measure of real-world clinical accuracy. Do not use for veterinary diagnosis.

Summary

• Architecture: YOLOv8n-cls (Ultralytics classification mode), fine-tuned from yolov8n-cls.pt
• Task: 4-class coarse classification of dog skin condition from a single image
• Classes: atopic_dermatitis, bacterial_pyoderma, fungal_malassezia, healthy
• Input size: 224×224
• Best val top-1 accuracy: 0.8523 (epoch 24 of 50)
• Final-epoch val top-1: 0.8295

Training

Setting	Value
Base weights	yolov8n-cls.pt (Ultralytics, AGPL-3.0)
Epochs	50
Batch size	32
Image size	224
Optimizer	auto
Hardware	1× Tesla T4 (Lightning AI Studio), ~2 min
Seed	0 (deterministic)

Full hyperparameters are in args.yaml. Per-epoch metrics are in results.csv; training curves in results.png; confusion matrices in confusion_matrix.png / confusion_matrix_normalized.png.

Data

• Source: yashmotiani/dogs-skin-disease-dataset (CC0)
• Size: 439 images total → 351 train / 88 val across 4 coarse classes
• Caveats: small, web-scraped, class-imbalanced, coarse labels that lump distinct presentations together. Not a curated clinical gold set.

Usage

from ultralytics import YOLO

model = YOLO("best.pt")          # or hf_hub_download the weights first
result = model.predict("dog_skin.jpg")[0]
print(result.names[result.probs.top1], float(result.probs.top1conf))

Files

• weights/best.pt — best checkpoint (epoch 24, top-1 0.8523) — use this for inference
• weights/last.pt — final checkpoint (epoch 50)
• args.yaml, results.csv, results.png, confusion_matrix*.png, *_batch*.jpg — training provenance

Limitations & intended use

Intended purpose is engineering signal for the PetCare triage pipeline, not medical use. Known limitations: tiny validation set (88 imgs), coarse labels, web-domain shift, no clinical validation, and single-image, single-view assumptions. A future version will use a curated gold eval set and a finer, clinically meaningful label taxonomy.

License

Inherits AGPL-3.0 from the Ultralytics YOLOv8 base weights.