dino.py

import torch
import numpy as np
from PIL import Image

# NOTE:
# We intentionally avoid importing `datasets` at module import time so the API can
# start even if the optional DINO dependencies are not installed locally.
from sklearn.metrics.pairwise import cosine_similarity


def load_dino(device):
    model = torch.hub.load(
        "facebookresearch/dinov2",
        "dinov2_vits14"
    )
    model.eval().to(device)
    return model


def build_embeddings(dino, transform, device):
    # Lazy import to keep DINO optional.
    from datasets import load_dataset

    dataset = load_dataset(
        "AdarshDS/mold-reference-images",
        split="train"
    )

    embs = []

    for sample in dataset:
        img: Image.Image = sample["image"].convert("RGB")
        t = transform(img).unsqueeze(0).to(device)

        with torch.no_grad():
            e = dino(t)

        embs.append(e.squeeze().cpu().numpy())

    if not embs:
        raise RuntimeError(
            "No reference images found in HF dataset"
        )

    return np.vstack(embs)


def similarity(dino, mold_embs, image, transform, device):
    t = transform(image).unsqueeze(0).to(device)
    with torch.no_grad():
        e = dino(t).cpu().numpy()

    return float(cosine_similarity(e, mold_embs).max())