fetch_artwork.py

"""
Fetches the 'artwork' domain (WikiArt-derived) images for NoOp-Bench.

WikiArt's own terms of use (https://www.wikiart.org/store/terms-of-use)
prohibit reproducing or republishing their content on other sites without
written consent, so these images are NOT bundled in noopbench_inputs_public.zip.
Run this script with your own Kaggle credentials to reconstruct them locally.

Requires: pip install kaggle Pillow
Requires: ~/.kaggle/kaggle.json with your own Kaggle API credentials
          (see https://www.kaggle.com/docs/api for how to generate one)

Usage: python fetch_artwork.py [--count 2500] [--out data_512_full/artwork]
"""

import argparse
import csv
import time
import zipfile
from pathlib import Path

from PIL import Image

parser = argparse.ArgumentParser()
parser.add_argument("--count", type=int, default=2500, help="number of images to fetch")
parser.add_argument("--out", type=str, default="data_512_full/artwork", help="output directory")
parser.add_argument("--pacing", type=float, default=1.5, help="seconds between requests")
args = parser.parse_args()

OUT = Path(args.out)
TMP = Path("_tmp_wikiart_download")
OUT.mkdir(parents=True, exist_ok=True)
TMP.mkdir(parents=True, exist_ok=True)

from kaggle import KaggleApi

api = KaggleApi()
api.authenticate()
print("Kaggle auth OK")


def resize_file(src, out_path):
    try:
        img = Image.open(src).convert("RGB")
        img = img.resize((512, 512), Image.LANCZOS)
        img.save(out_path)
        return True
    except Exception as e:
        print(f"  [skip] {out_path.name}: {e}")
        return False


def download_kaggle_file(dataset_slug, file_name, dest_dir, max_retries=4, backoff_base=20):
    # Nested paths (e.g. "Style/artist_title.jpg") need the "/" percent-encoded,
    # otherwise Kaggle's API routes it as a literal path segment and 404s.
    api_name = file_name.replace("/", "%2F")
    for attempt in range(max_retries):
        try:
            api.dataset_download_file(dataset_slug, api_name, path=str(dest_dir), force=False, quiet=True)
            return True
        except Exception as e:
            if "429" in str(e) and attempt < max_retries - 1:
                wait = backoff_base * (attempt + 1)
                print(f"  [429] backing off {wait}s...")
                time.sleep(wait)
                continue
            return False
    return False


def fetch_one(fname, out_idx):
    out = OUT / f"artwork_wikiart_{out_idx:05d}.png"
    if out.exists():
        return True
    if not download_kaggle_file("steubk/wikiart", fname, TMP):
        return False
    basename = Path(fname).name
    # Kaggle saves nested-path files with the encoded path baked into the
    # filename (e.g. "Style%2Fartist_title.jpg"), not a clean basename.
    downloaded = [p for p in TMP.rglob("*") if p.is_file() and p.name.endswith(basename)]
    if downloaded:
        ok = resize_file(downloaded[0], out)
        downloaded[0].unlink(missing_ok=True)
        return ok
    return False


existing = len(list(OUT.glob("*.png")))
needed = args.count
print(f"artwork: {existing}/{needed} already have")

if existing < needed:
    img_files = []
    page_token = None
    while len(img_files) < needed + 200:
        resp = api.dataset_list_files("steubk/wikiart", page_token=page_token, page_size=200)
        batch = [f.name for f in resp.dataset_files if f.name.lower().endswith((".jpg", ".jpeg", ".png"))]
        img_files.extend(batch)
        page_token = getattr(resp, "next_page_token", None)
        if not page_token or not resp.dataset_files:
            break
    print(f"listed {len(img_files)} candidate files")

    manifest_path = Path("artwork_fetch_manifest.csv")
    with open(manifest_path, "a", newline="") as mf:
        writer = csv.writer(mf)
        if manifest_path.stat().st_size == 0:
            writer.writerow(["out_index", "source_filename"])

        done = existing
        errors = 0
        for n, fname in enumerate(img_files[existing:]):
            if done >= needed:
                break
            out_idx = existing + n
            ok = fetch_one(fname, out_idx)
            if ok:
                writer.writerow([out_idx, fname])
                mf.flush()
                done += 1
            else:
                errors += 1
            if done % 50 == 0:
                print(f"  artwork: {done}/{needed} (errors so far: {errors})")
            time.sleep(args.pacing)

print(f"artwork done: {len(list(OUT.glob('*.png')))} images")
print(f"Manifest of source filenames written to artwork_fetch_manifest.csv")