Base model: google/vit-base-patch16-224
Architecture: ViT-Base (86.6M params, patch size 16, input 224×224)
Task: Multi-class image classification (25 classes)
Training data: meganariley/inat-wildflowers-25 — 10,000 research-grade iNaturalist photos

Performance

Split	Accuracy	Loss
Validation	97.1%	0.098
Test	97.9%	0.082

Per-Class Test Accuracy

Species	Accuracy	Species	Accuracy
Blazing Star	95.0%	Marsh Marigold	97.5%
Blue Flag Iris	100.0%	Ox-eye Daisy	95.0%
California Poppy	97.5%	Queen Anne's Lace	95.0%
Canada Goldenrod	100.0%	Spring Beauty	100.0%
Cardinal Flower	100.0%	Trout Lily	100.0%
Colorado Columbine	97.5%	Virginia Bluebells	97.5%
Common Milkweed	97.5%	White Trillium	97.5%
Common Yarrow	100.0%	Wild Bergamot	97.5%
Dutchman's Breeches	100.0%	Wild Columbine	100.0%
Fireweed	95.0%	Wild Lupine	100.0%
Indian Blanket	97.5%	Wild Rose	100.0%
Indian Paintbrush	95.0%	Joe-Pye Weed	95.0%
Jack-in-the-Pulpit	97.5%

10 species at 100% accuracy · All species ≥ 95%

Training

Epochs: 5
Optimizer: AdamW (lr=5e-5, weight_decay=0.01)
Batch size: 16 × 4 gradient accumulation = effective 64
Warmup: 10% linear
Augmentation: RandomResizedCrop, HorizontalFlip, ColorJitter
Hardware: CPU (8 vCPU / 32GB RAM)
Training time: ~2 hours

Training Curve

Epoch	Val Accuracy	Val Loss
1	93.5%	0.255
2	95.2%	0.146
3	96.3%	0.114
4	97.0%	0.104
5	97.1%	0.098

Usage

python

from transformers import pipeline

classifier = pipeline("image-classification", model="meganariley/wildflower-classifier-vit")
results = classifier("path/to/wildflower.jpg", top_k=5)
for r in results:
    print(f"{r['label']}: {r['score']:.3f}")

Demo

Try it live: 🌸 Wildflower Identifier Space

License

Apache 2.0 (base model license). Training data from iNaturalist under CC BY-NC 4.0.