scripts/cloud/embed_modal.py

"""Modal GPU job for CLIP embedding — streams images from HF, no zip extraction.

Streams images directly from AI4Patents/IMPACT using the HF datasets API
(record-by-record, no zip download). Runs CLIP ViT-L/14 on GPU.
Pushes embeddings parquet to HF Hub immediately — survives any /tmp cleanup.

Setup (one time):
    modal setup
    modal secret create hf-secret HF_TOKEN=hf_...

Run:
    modal run scripts/cloud/embed_modal.py --year 2022

Output: hf://datasets/midah/patent-wireframes/embeddings/{year}_vitl14.parquet
"""

import io
import os

import modal

image = (
    modal.Image.debian_slim(python_version="3.11")
    .pip_install(
        "open_clip_torch",
        "huggingface_hub",
        "datasets",
        "Pillow",
        "pandas",
        "numpy",
        "tqdm",
        "requests",
    )
)

import time as _time
app = modal.App(f"patent-clip-{int(_time.time()) % 10000}", image=image)
hf_secret = modal.Secret.from_name("hf-secret")

OUT_REPO = "midah/patent-wireframes"


@app.function(
    gpu="T4",       # T4 less preempted than A10G; slower but completes
    timeout=21600,  # 6 hours — T4 is slower, full scan ~5.5 hrs
    secrets=[hf_secret],
    memory=32768,
    retries=modal.Retries(max_retries=3, backoff_coefficient=2.0, initial_delay=10.0),
)
def embed_year(year: str = "2022", model_name: str = "ViT-L-14",
               pretrained: str = "openai", batch_size: int = 64):
    """Stream images from IMPACT HF dataset, embed with CLIP, push to Hub."""
    import ast
    import base64
    import csv
    from pathlib import Path

    import numpy as np
    import open_clip
    import pandas as pd
    import requests
    import torch
    from huggingface_hub import HfApi, hf_hub_download
    from PIL import Image
    from tqdm import tqdm

    token = os.environ["HF_TOKEN"]
    device = "cuda"

    # ── Load CLIP ─────────────────────────────────────────────────────────────
    print(f"Loading {model_name} ({pretrained}) on {device}...")
    model, _, preprocess = open_clip.create_model_and_transforms(
        model_name, pretrained=pretrained
    )
    model = model.to(device).eval()
    print("Model loaded.")

    # ── Download zip and metadata CSV ────────────────────────────────────────
    print(f"Downloading IMPACT {year} zip (~4.4GB) and CSV...")
    zip_path = hf_hub_download(
        repo_id="AI4Patents/IMPACT",
        filename=f"{year}.zip",
        repo_type="dataset",
        token=token,
        local_dir="/tmp/impact",
    )
    csv_path = hf_hub_download(
        repo_id="AI4Patents/IMPACT",
        filename=f"{year}.csv",
        repo_type="dataset",
        token=token,
        local_dir="/tmp/impact",
    )

    # Build figure list from CSV to get patent IDs for output
    pid_map = {}  # filename → (patent_id, figure_num)
    with open(csv_path) as f:
        for row in csv.DictReader(f):
            try:
                fnames = ast.literal_eval(row.get("file_names") or "[]")
                pid = row.get("id") or ""
                for i, fn in enumerate(fnames):
                    pid_map[fn] = (pid, i)
            except Exception:
                pass

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

    # ── Sequential mmap scan of zip — avoids EOCD / central directory ────────
    # The zip has a corrupt/non-standard Zip64 central directory that defeats
    # Python's zipfile and 7z extraction. Instead, scan local file headers
    # (PK\x03\x04) sequentially using mmap — no central directory needed.
    # Each TIF is decompressed in memory and fed directly to GPU.

    import mmap, zlib, struct

    SIG_LOCAL = b"PK\x03\x04"

    def scan_zip_and_embed(zip_path: str):
        """Yield (filename, PIL.Image) by scanning zip local headers."""
        with open(zip_path, "rb") as f:
            mm = mmap.mmap(f.fileno(), 0, access=mmap.ACCESS_READ)
            pos = 0
            total_size = mm.size()
            while pos < total_size - 30:
                idx = mm.find(SIG_LOCAL, pos)
                if idx < 0:
                    break
                header = mm[idx: idx + 30]
                if len(header) < 30:
                    break
                flags       = struct.unpack_from("<H", header, 6)[0]
                comp_method = struct.unpack_from("<H", header, 8)[0]
                comp_size   = struct.unpack_from("<I", header, 18)[0]
                fname_len   = struct.unpack_from("<H", header, 26)[0]
                extra_len   = struct.unpack_from("<H", header, 28)[0]
                has_dd      = bool(flags & 0x0008)

                fname_bytes = mm[idx + 30: idx + 30 + fname_len]
                fname = fname_bytes.decode("utf-8", errors="replace")
                data_start = idx + 30 + fname_len + extra_len

                if not fname.endswith(".TIF"):
                    pos = data_start + max(comp_size, 1) if not has_dd else data_start + 1
                    continue

                # For data-descriptor files, find the next local header to bound data
                if has_dd or comp_size == 0:
                    next_sig = mm.find(SIG_LOCAL, data_start + 4)
                    if next_sig < 0:
                        next_sig = total_size
                    # Data ends 16 bytes before next sig (data descriptor)
                    raw_data = mm[data_start: next_sig - 16]
                    pos = next_sig
                else:
                    raw_data = mm[data_start: data_start + comp_size]
                    pos = data_start + comp_size

                try:
                    if comp_method == 8:  # deflate
                        img_bytes = zlib.decompress(raw_data, -15)
                    elif comp_method == 0:  # stored
                        img_bytes = bytes(raw_data)
                    else:
                        continue
                    basename = fname.split("/")[-1]
                    img = Image.open(io.BytesIO(img_bytes)).convert("RGB")
                    yield basename, img
                except Exception:
                    continue
            mm.close()

    def fetch_image(fig: dict) -> Image.Image | None:
        return None  # unused — we use scan_zip_and_embed instead

    # ── Resume from existing HF checkpoint ───────────────────────────────────
    # Download any existing embeddings from HF so we accumulate across preemptions
    # rather than restarting from zero each time.
    all_ids: list[str] = []
    all_vecs: list[np.ndarray] = []
    already_embedded: set[str] = set()

    existing_path = f"/tmp/{year}_existing.parquet"
    try:
        existing_hf = hf_hub_download(
            repo_id=OUT_REPO, filename=f"embeddings/{year}_vitl14.parquet",
            repo_type="dataset", token=token, local_dir="/tmp/impact",
        )
        existing_df = pd.read_parquet(existing_hf)
        already_embedded = set(existing_df["figure_id"].tolist())
        # Seed all_ids / all_vecs with existing embeddings
        all_ids = existing_df["figure_id"].tolist()
        # Convert object-array of 768-dim vectors to proper 2D float32 array
        existing_vecs = np.vstack(existing_df["embedding"].tolist()).astype(np.float32)
        all_vecs = [existing_vecs]
        print(f"Resuming from {len(already_embedded):,} existing embeddings on HF")
    except Exception:
        print("No existing checkpoint — starting fresh")

    batch_imgs: list[Image.Image] = []
    batch_ids: list[str] = []
    n_processed = 0
    last_checkpoint_n = len(all_ids) // 10000  # start above existing checkpoints
    CHECKPOINT_EVERY = 25000  # unused field — kept for reference

    def flush_batch():
        if not batch_imgs:
            return
        tensors = torch.stack([preprocess(im) for im in batch_imgs]).to(device)
        with torch.no_grad():
            feats = model.encode_image(tensors)
            feats = feats / feats.norm(dim=-1, keepdim=True)
        all_vecs.append(feats.cpu().numpy())
        all_ids.extend(batch_ids)
        batch_imgs.clear()
        batch_ids.clear()

    print(f"Scanning zip and embedding (sequential mmap, no EOCD needed)...")
    for basename, img in scan_zip_and_embed(zip_path):
        n_processed += 1
        if n_processed % 5000 == 0:
            print(f"  {n_processed:,} files scanned, {len(all_ids):,} embedded")

        # Map basename back to patent ID using pid_map
        pid_raw, fig_num = pid_map.get(basename, (None, None))
        if pid_raw is None:
            continue

        pid_norm = str(pid_raw).lstrip("D").zfill(7)
        fig_id = f"D{pid_norm}_{fig_num}"

        # Skip already-embedded figures
        if fig_id in already_embedded:
            continue

        batch_ids.append(fig_id)
        batch_imgs.append(img)
        if len(batch_imgs) >= batch_size:
            flush_batch()

        # Checkpoint every 10k NEW embeddings to HF
        current_boundary = len(all_ids) // 10000
        if current_boundary > last_checkpoint_n and len(all_ids) > len(already_embedded):
            _vecs_cp = np.vstack(all_vecs).astype(np.float32)
            _norms_cp = np.linalg.norm(_vecs_cp, axis=1, keepdims=True)
            _vecs_cp /= np.maximum(_norms_cp, 1e-8)
            _df_cp = pd.DataFrame({"figure_id": list(all_ids), "embedding": list(_vecs_cp)})
            _out_cp = f"/tmp/{year}_cp.parquet"
            _df_cp.to_parquet(_out_cp, index=False)
            HfApi(token=token).upload_file(
                path_or_fileobj=_out_cp,
                path_in_repo=f"embeddings/{year}_vitl14.parquet",
                repo_id=OUT_REPO, repo_type="dataset",
                commit_message=f"Checkpoint {len(all_ids):,} for {year}",
            )
            print(f"  Checkpoint: {len(all_ids):,} embeddings on HF")
            last_checkpoint_n = current_boundary

    flush_batch()
    print(f"Scan complete: {n_processed:,} files scanned, {len(all_ids):,} embedded")

    print(f"Embedded: {len(all_ids):,} figures")
    if not all_ids:
        print("No figures embedded — check HF access")
        return {"n": 0}

    # ── Normalize + save + push ───────────────────────────────────────────────
    vecs = np.vstack(all_vecs).astype(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 = f"/tmp/{year}_vitl14.parquet"
    df.to_parquet(local_out, index=False)
    size_mb = Path(local_out).stat().st_size / 1e6
    print(f"Parquet: {size_mb:.1f}MB — pushing to HF Hub...")

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

    return {"year": year, "n_embedded": len(all_ids), "shape": list(vecs.shape)}


@app.local_entrypoint()
def main(year: str = "2022"):
    print(f"Embedding year: {year}")
    result = embed_year.remote(year)
    print("Done:", result)