src/backends/hf_backend.py

import numpy as np
import torch
from PIL import Image
from transformers import CLIPModel, CLIPProcessor, ViTFeatureExtractor, ViTForImageClassification

_clip_model = None
_clip_processor = None
_age_model = None
_age_extractor = None

# nateraw/vit-age-classifier outputs probabilities over these age range buckets
_AGE_LABELS = ["0-2", "3-9", "10-19", "20-29", "30-39", "40-49", "50-59", "60-69", "70+"]
_AGE_MIDPOINTS = [1, 6, 14, 24, 34, 44, 54, 64, 74]


def _get_clip():
    global _clip_model, _clip_processor
    if _clip_model is None:
        _clip_model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
        _clip_processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
    return _clip_model, _clip_processor


def _get_age_model():
    global _age_model, _age_extractor
    if _age_model is None:
        _age_model = ViTForImageClassification.from_pretrained("nateraw/vit-age-classifier")
        _age_extractor = ViTFeatureExtractor.from_pretrained("nateraw/vit-age-classifier")
    return _age_model, _age_extractor


def _embed(image_path):
    model, processor = _get_clip()
    image = Image.open(image_path).convert("RGB")
    inputs = processor(images=image, return_tensors="pt")
    with torch.no_grad():
        features = model.get_image_features(**inputs)
    features = features / features.norm(dim=-1, keepdim=True)
    return features[0].numpy()


def _estimate_age(image_path):
    model, extractor = _get_age_model()
    image = Image.open(image_path).convert("RGB")
    inputs = extractor(images=image, return_tensors="pt")
    with torch.no_grad():
        outputs = model(**inputs)
    probs = torch.softmax(outputs.logits, dim=-1)[0].numpy()
    return int(round(float(np.dot(probs, _AGE_MIDPOINTS))))


def _to_pct(sim):
    # CLIP features are already L2-normalized so dot product == cosine similarity (-1 to 1)
    return round(max(0.0, min(1.0, (float(sim) + 1) / 2)) * 100, 1)


def analyze(father_path, mother_path, child_path):
    child_emb = _embed(child_path)
    father_emb = _embed(father_path)
    mother_emb = _embed(mother_path)

    return {
        "age": _estimate_age(child_path),
        "father_score": _to_pct(np.dot(child_emb, father_emb)),
        "mother_score": _to_pct(np.dot(child_emb, mother_emb)),
    }