scripts/eval/embed_figures.py

"""Compute CLIP embeddings for all extracted patent figures.

Uses open_clip ViT-L/14 (best open-source CLIP) to embed each TIF.
Saves embeddings as parquet with figure_id index for fast lookup.

Usage:
    python scripts/eval/embed_figures.py \
        --enriched data/enriched/enriched_2022.parquet \
        --images /tmp/patent_sample/2022 \
        --out data/embeddings/embeddings_2022.parquet \
        --batch 64
"""

import argparse
from pathlib import Path

import numpy as np
import open_clip
import pandas as pd
import torch
from PIL import Image
from tqdm import tqdm


def find_image(images_dir: Path, image_filename: str) -> Path | None:
    parts = image_filename.split("-D0")
    if len(parts) < 2:
        return None
    p = images_dir / parts[0] / image_filename
    return p if p.exists() else None


def embed_batch(model, preprocess, paths: list[Path], device: str) -> np.ndarray:
    imgs = []
    for p in paths:
        try:
            img = Image.open(p).convert("RGB")
            imgs.append(preprocess(img))
        except Exception:
            imgs.append(torch.zeros(3, 224, 224))
    batch = torch.stack(imgs).to(device)
    with torch.no_grad():
        feats = model.encode_image(batch)
        feats = feats / feats.norm(dim=-1, keepdim=True)
    return feats.cpu().numpy()


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--enriched", default="data/enriched/enriched_2022.parquet")
    parser.add_argument("--images", default="/tmp/patent_sample/2022")
    parser.add_argument("--out", default="data/embeddings/embeddings_2022.parquet")
    parser.add_argument("--batch", type=int, default=64)
    parser.add_argument("--model", default="ViT-L-14")
    parser.add_argument("--pretrained", default="openai")
    args = parser.parse_args()

    device = "cuda" if torch.cuda.is_available() else "cpu"
    print(f"Device: {device}")

    print(f"Loading {args.model} ({args.pretrained})...")
    model, _, preprocess = open_clip.create_model_and_transforms(
        args.model, pretrained=args.pretrained
    )
    model = model.to(device).eval()

    df = pd.read_parquet(args.enriched)
    images_dir = Path(args.images)

    # Resolve image paths
    df["_img_path"] = df["image_filename"].apply(
        lambda fn: find_image(images_dir, fn)
    )
    found = df["_img_path"].notna().sum()
    print(f"Images resolved: {found:,} / {len(df):,}")
    df = df[df["_img_path"].notna()].reset_index(drop=True)

    # Embed in batches
    all_ids, all_vecs = [], []
    batch_size = args.batch
    paths = df["_img_path"].tolist()
    fig_ids = df["figure_id"].tolist()

    for i in tqdm(range(0, len(paths), batch_size), desc="Embedding"):
        batch_paths = paths[i : i + batch_size]
        batch_ids = fig_ids[i : i + batch_size]
        vecs = embed_batch(model, preprocess, batch_paths, device)
        all_ids.extend(batch_ids)
        all_vecs.append(vecs)

    all_vecs = np.vstack(all_vecs)
    print(f"Embedding matrix: {all_vecs.shape}")

    Path(args.out).parent.mkdir(parents=True, exist_ok=True)
    out_df = pd.DataFrame({"figure_id": all_ids})
    # Store each embedding dimension as a column — efficient for FAISS loading
    out_df["embedding"] = list(all_vecs)
    out_df.to_parquet(args.out, index=False)
    print(f"Saved → {args.out}")

    # Quick sanity check
    print(f"Sample figure: {all_ids[0]}")
    print(f"Embedding norm: {np.linalg.norm(all_vecs[0]):.4f} (should be 1.0)")


if __name__ == "__main__":
    main()