juppy44/plant-identification-2m-vit-b

Highlights

• Base model: google/vit-base-patch16-224
• Parameters: ~97.2M
• Training samples: 2,000,000 curated plant occurrences
• Species coverage: ~14,000 unique species
• Source data: GBIF (research-grade iNaturalist images)
• Training method: End-to-end supervised fine-tuning
• Use-case: Fast species-level classification from a single photo

---

Example Usage

from transformers import AutoImageProcessor, AutoModelForImageClassification
from PIL import Image
import requests
import torch

model_id = "juppy44/plant-identification-2m-vit-b"

processor = AutoImageProcessor.from_pretrained(model_id)
model = AutoModelForImageClassification.from_pretrained(model_id)

url = "https://example.com/plant.jpg"
image = Image.open(requests.get(url, stream=True).raw)

inputs = processor(images=image, return_tensors="pt")
with torch.no_grad():
    logits = model(**inputs).logits

pred = logits.softmax(dim=-1)[0]
topk = torch.topk(pred, k=5)

for prob, idx in zip(topk.values, topk.indices):
    label = model.config.id2label[idx.item()]
    print(f"{label}: {prob.item():.4f}")

---

Intended Applications

• Ecological surveys
• Nursery and horticulture tools
• Restoration and revegetation workflows
• Field research and biodiversity monitoring
• Citizen science and educational platforms
• Image-based species tagging pipelines

---

Data & Training Details

Dataset Construction

• Sourced from GBIF occurrences with valid species and image metadata.
• Cleaned and deduplicated.
• Species filtered to those with ≥ 20 images.
• Maximum cap of 1,000 images per species to reduce class imbalance.
• Final training dataset: 2,000,000 images across ~14k species.

Training

• ViT-Base fine-tuned for 1 epoch over 2M samples.
• AdamW optimizer, standard ViT augmentations.
• Mixed-precision training on GPU.

Limitations

• Some species are visually indistinguishable without context (location, reproduction structures, etc.).
• Performance varies for rare, morphologically similar, or poorly photographed species.
• No location metadata incorporated yet — purely image-based.

---

Labels

Species names follow the canonical GBIF taxonomy (species_name). Each class corresponds directly to a species.

You can inspect all labels via:

from transformers import AutoConfig
cfg = AutoConfig.from_pretrained("juppy44/plant-identification-2m-vit-b")
labels = cfg.id2label

---

Performance

No fixed metric is published yet due to absence of a clean evaluation split in GBIF/iNat. However, the model performs strongly in practical testing and is suitable for downstream adaptation or LoRA-based fine-tuning.

Formal benchmarks will be added when a standardized evaluation subset is released.

---

Fine-Tuning & Adapters

You can further specialize the model using LoRA adapters for:

• Regional subsets
• Functional groups
• Threatened species
• Agricultural crops
• Disease classification

The base model is trained broadly enough to support domain-specific adapter tuning with minimal compute.

---

License

Model weights follow the same license as the underlying ViT-Base model. Users are responsible for ensuring compliance with GBIF/iNaturalist usage terms for any downstream dataset creation.

---