Feline Dermatology Classifier — Howl Vision
EfficientNetV2-S for feline skin lesion classification (4 classes). Part of Howl Vision for the Gemma 4 Good Hackathon.
Metrics (held-out test, n=152)
| Metric | Value |
|---|---|
| Accuracy | 90.1% [84.4%, 93.9%] Wilson CI 95% |
| F1-macro | 0.902 |
| Cohen's Kappa | 0.868 |
Per-class: Flea_Allergy 0.904, Health 0.923, Ringworm 0.861, Scabies 0.919
Details
- Architecture:
timm/tf_efficientnetv2_s.in21k_ft_in1k, num_classes=4 - Input: 384x384 RGB, ImageNet normalization
- Classes: Flea_Allergy, Health, Ringworm, Scabies
- Training: 699 images, AdamW lr=1e-4, CosineAnnealing, early stop epoch 16/30
import timm, torch
model = timm.create_model("tf_efficientnetv2_s.in21k_ft_in1k", pretrained=False, num_classes=4)
ckpt = torch.load("vet_feline_dermatology.pt", map_location="cpu", weights_only=False)
model.load_state_dict(ckpt["model_state_dict"])
model.eval()