scripts/cloud/create_hf_endpoint.py

"""Create, use, and delete an HF Inference Endpoint for CLIP embedding.

Creates a temporary dedicated endpoint for batch CLIP embedding,
processes all images from the IMPACT dataset, pushes embeddings
to HF Hub, then deletes the endpoint.

Cost estimate:
  DesignCLIP / ViT-L-14 on T4: ~$0.08/hr
  103k images (2022) at batch=64: ~25 min = ~$0.03
  3.61M images (all years) at batch=64: ~15 hrs = ~$1.20 total

Usage:
    export HF_TOKEN=hf_...
    python scripts/cloud/create_hf_endpoint.py \
        --year 2022 \
        --model openai/clip-vit-large-patch14 \
        --instance-type aws-us-east-1-t4g-small \
        --out-repo midah/patent-wireframes
"""

import argparse
import base64
import io
import os
import time
from pathlib import Path

import requests
from dotenv import load_dotenv

load_dotenv(".env")

HF_ENDPOINTS_API = "https://api.endpoints.huggingface.cloud/v2/endpoint"
HF_TOKEN = os.environ.get("HF_TOKEN")
HEADERS = {"Authorization": f"Bearer {HF_TOKEN}", "Content-Type": "application/json"}


# ── Endpoint lifecycle ────────────────────────────────────────────────────────

def create_endpoint(name: str, model: str, instance_type: str) -> dict:
    """Create a dedicated inference endpoint for the given model."""
    payload = {
        "accountId": None,
        "compute": {
            "accelerator": "gpu" if "t4" in instance_type or "a10g" in instance_type else "cpu",
            "instanceSize": instance_type.split("-")[-1] if "-" in instance_type else "large",
            "instanceType": instance_type,
            "scaling": {"maxReplica": 1, "minReplica": 1},
        },
        "model": {
            "framework": "pytorch",
            "image": {"huggingface": {"env": {}}},
            "repository": model,
            "revision": "main",
            "task": "feature-extraction",
        },
        "name": name,
        "provider": {"region": "us-east-1", "vendor": "aws"},
        "type": "protected",
    }
    r = requests.post(
        f"{HF_ENDPOINTS_API}/midah",
        headers=HEADERS,
        json=payload,
        timeout=30,
    )
    r.raise_for_status()
    return r.json()


def wait_for_endpoint(name: str, timeout: int = 600) -> str:
    """Poll until the endpoint is running. Returns the endpoint URL."""
    start = time.time()
    while time.time() - start < timeout:
        r = requests.get(f"{HF_ENDPOINTS_API}/midah/{name}", headers=HEADERS, timeout=15)
        r.raise_for_status()
        data = r.json()
        state = data.get("status", {}).get("state", "unknown")
        url = data.get("status", {}).get("url", "")
        print(f"  State: {state} ({int(time.time()-start)}s elapsed)")
        if state == "running":
            return url
        if state in ("failed", "scaledToZero"):
            raise RuntimeError(f"Endpoint failed: {state}")
        time.sleep(20)
    raise TimeoutError("Endpoint did not start within timeout")


def delete_endpoint(name: str):
    r = requests.delete(f"{HF_ENDPOINTS_API}/midah/{name}", headers=HEADERS, timeout=15)
    if r.status_code not in (200, 204):
        print(f"  Warning: delete returned {r.status_code}")
    else:
        print(f"  Deleted endpoint: {name}")


# ── Embedding via endpoint ────────────────────────────────────────────────────

def encode_image(img_bytes: bytes, max_edge: int = 224) -> str:
    from PIL import Image
    img = Image.open(io.BytesIO(img_bytes)).convert("RGB")
    w, h = img.size
    scale = min(max_edge / max(w, h), 1.0)
    if scale < 1.0:
        img = img.resize((int(w * scale), int(h * scale)), Image.LANCZOS)
    buf = io.BytesIO()
    img.save(buf, format="JPEG", quality=85)
    return base64.standard_b64encode(buf.getvalue()).decode()


def embed_batch(endpoint_url: str, b64_images: list[str]) -> list[list[float]] | None:
    """Call the endpoint with a batch of base64 images."""
    payload = {"inputs": b64_images}
    for attempt in range(4):
        try:
            r = requests.post(
                endpoint_url,
                headers={**HEADERS, "Content-Type": "application/json"},
                json=payload,
                timeout=60,
            )
            if r.status_code == 200:
                data = r.json()
                # Response shape: list of embeddings or list of list of list
                if isinstance(data, list) and data:
                    if isinstance(data[0], list) and isinstance(data[0][0], float):
                        return data  # already [[float, ...], ...]
                    if isinstance(data[0], list) and isinstance(data[0][0], list):
                        return [d[0] for d in data]  # [[[float]], ...]
                return None
            elif r.status_code in (429, 503):
                time.sleep(2 ** attempt)
        except Exception as e:
            print(f"  Batch error (attempt {attempt+1}): {e}")
            time.sleep(2 ** attempt)
    return None


# ── Main processing ───────────────────────────────────────────────────────────

def process_year(year: str, endpoint_url: str, out_repo: str, batch_size: int = 32):
    """Stream images from IMPACT and embed via the endpoint."""
    import ast
    import csv
    import zipfile

    import numpy as np
    import pandas as pd
    from huggingface_hub import HfApi, hf_hub_download

    token = HF_TOKEN
    api = HfApi(token=token)

    print(f"\nDownloading IMPACT {year} CSV...")
    csv_path = hf_hub_download(
        repo_id="AI4Patents/IMPACT", filename=f"{year}.csv",
        repo_type="dataset", token=token,
    )

    print(f"Downloading IMPACT {year} images zip (~4.4GB)...")
    zip_path = hf_hub_download(
        repo_id="AI4Patents/IMPACT", filename=f"{year}.zip",
        repo_type="dataset", token=token,
    )

    # Build figure list
    figures = []
    with open(csv_path) as f:
        for row in csv.DictReader(f):
            try:
                fnames = ast.literal_eval(row["file_names"])
                pid = row["id"]
                for i, fn in enumerate(fnames):
                    figures.append({"patent_id": pid, "figure_num": i, "filename": fn})
            except Exception:
                pass

    print(f"Total figures: {len(figures):,}")

    all_ids, all_vecs = [], []
    n_failed = 0

    # Process in batches using zip
    with zipfile.ZipFile(zip_path) as zf:
        batch_imgs, batch_ids = [], []

        def flush():
            nonlocal n_failed
            if not batch_imgs:
                return
            vecs = embed_batch(endpoint_url, batch_imgs)
            if vecs:
                all_vecs.extend(vecs)
                all_ids.extend(batch_ids)
            else:
                n_failed += len(batch_imgs)
            batch_imgs.clear()
            batch_ids.clear()

        from tqdm import tqdm
        for fig in tqdm(figures, desc=f"Embedding {year}"):
            fn = fig["filename"]
            parts = fn.split("-D0")
            if len(parts) < 2:
                continue
            inner = f"{year}/{parts[0]}/{fn}"
            try:
                with zf.open(inner) as f:
                    img_bytes = f.read()
                b64 = encode_image(img_bytes)
                pid = fig["patent_id"].lstrip("D").zfill(7)
                batch_ids.append(f"D{pid}_{fig['figure_num']}")
                batch_imgs.append(b64)
                if len(batch_imgs) >= batch_size:
                    flush()
            except Exception:
                continue

        flush()

    print(f"Embedded: {len(all_ids):,} | Failed: {n_failed}")

    if not all_ids:
        print("No embeddings produced — check endpoint")
        return

    # Normalize and save
    vecs = np.array(all_vecs, dtype=np.float32)
    norms = np.linalg.norm(vecs, axis=1, keepdims=True)
    vecs /= np.maximum(norms, 1e-8)

    df = pd.DataFrame({"figure_id": all_ids, "embedding": list(vecs)})
    out_file = f"embeddings_{year}_vitl14.parquet"
    local_out = Path(f"/tmp/{out_file}")
    df.to_parquet(local_out, index=False)
    size_mb = local_out.stat().st_size / 1e6
    print(f"Parquet: {size_mb:.1f}MB")

    api.upload_file(
        path_or_fileobj=str(local_out),
        path_in_repo=f"embeddings/{out_file}",
        repo_id=out_repo,
        repo_type="dataset",
        commit_message=f"Add CLIP embeddings for {year}",
    )
    print(f"Pushed → hf://datasets/{out_repo}/embeddings/{out_file}")
    local_out.unlink()


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--year",          default="2022")
    parser.add_argument("--model",         default="openai/clip-vit-large-patch14")
    parser.add_argument("--instance-type", default="aws-us-east-1-t4g-small",
                        help="HF endpoint instance type. See: hf.co/docs/inference-endpoints")
    parser.add_argument("--out-repo",      default="midah/patent-wireframes")
    parser.add_argument("--batch",         type=int, default=32)
    parser.add_argument("--keep-endpoint", action="store_true",
                        help="Don't delete endpoint after finishing (useful for multi-year runs)")
    args = parser.parse_args()

    if not HF_TOKEN:
        raise RuntimeError("HF_TOKEN not set")

    endpoint_name = f"patent-clip-{args.year}-{int(time.time())}"
    endpoint_url = None

    try:
        print(f"Creating endpoint: {endpoint_name}")
        print(f"  Model: {args.model}")
        print(f"  Instance: {args.instance_type}")
        data = create_endpoint(endpoint_name, args.model, args.instance_type)
        print(f"  Created. Waiting for running state...")
        endpoint_url = wait_for_endpoint(endpoint_name)
        print(f"  Endpoint ready: {endpoint_url}")

        process_year(args.year, endpoint_url, args.out_repo, args.batch)

    finally:
        if not args.keep_endpoint and endpoint_name:
            print(f"\nCleaning up endpoint: {endpoint_name}")
            delete_endpoint(endpoint_name)


if __name__ == "__main__":
    main()