envision-classifier 0.2.0

Few-shot classifier for detecting eye imaging datasets

envision-classifier

SetFit few-shot classifier for identifying eye imaging datasets from scientific metadata.

Part of the EyeACT project by the FAIR Data Innovations Hub.

Installation

pip install envision-classifier

Python API

from envision_classifier import EyeImagingClassifier

# Downloads model from HuggingFace on first use
clf = EyeImagingClassifier()

# Classify a single record
result = clf.classify("Retinal OCT dataset for diabetic retinopathy")
print(result)
# {'label': 'EYE_IMAGING', 'confidence': 0.98,
#  'probabilities': {'EYE_IMAGING': 0.98, 'NEGATIVE': 0.02}}

# Classify a batch
results = clf.classify_batch([
    "Retinal fundus photography dataset for glaucoma screening",
    "COVID-19 genome sequencing data",
    {"title": "OCT images", "description": "Macular degeneration scans"},
])

# Use a local model instead of downloading
clf = EyeImagingClassifier(model_path="./my_model")

CLI

After installing, the envision-classifier command is available:

# Classify a text string
envision-classifier classify --text "Retinal OCT dataset for diabetic retinopathy"

# Classify from a JSON file
envision-classifier classify records.json

# Pipe JSON via stdin
echo '{"title": "Fundus images", "description": "DR screening"}' | envision-classifier classify

# Train a new model from built-in training data
envision-classifier train --output ./my_model

# Show model info and training data counts
envision-classifier info

Classification Labels

Label	Description
EYE_IMAGING	Actual eye imaging datasets (fundus, OCT, OCTA, cornea)
NEGATIVE	Everything else (software, non-imaging eye data, unrelated domains)

Model

• Base model: sentence-transformers/all-mpnet-base-v2 (768-dim)
• Training data: 891 curated examples (262 EYE_IMAGING, 629 NEGATIVE) from Zenodo, Figshare, Dryad, Kaggle, and NEI
• Test accuracy: 0.961, EYE_IMAGING F1: 0.936
• Spot-check: 30/33 (90.9%)
• Model weights: fairdataihub/envision-eye-imaging-classifier

Zenodo Classification Results

Applied to 515 Zenodo dataset records via envision-discovery:

Class	Count
EYE_IMAGING	60
NEGATIVE	455

Classification is based on metadata only (titles, descriptions, keywords, and file types inspected inside archives via HTTP Range requests) -- no dataset files are downloaded.

Related

• envision-discovery -- Full pipeline (scraping + classification + export)
• Model on HuggingFace

License

MIT