Add paginated thumbnail catalog for 1,400 SimReady assets

README.md

@@ -0,0 +1,36 @@
+---
+pretty_name: NVIDIA SimReady Assets
+tags:
+- 3d
+- simready
+configs:
+- config_name: catalog
+  default: true
+  data_files:
+  - split: assets
+    path: catalog/assets.parquet
+dataset_info:
+- config_name: catalog
+  features:
+  - name: thumbnail
+    dtype: image
+  - name: name
+    dtype: string
+  - name: asset_url
+    dtype: string
+---
+
+# NVIDIA SimReady Assets
+
+Browse **1,400 assets** in the thumbnail table on this page. Each row shows the
+asset's existing preview, its package name, and a link to its main USD file.
+
+[Open the full asset table](https://huggingface.co/datasets/nvidia/simready-assets/viewer/catalog/assets)
+to search by name and navigate between pages.
+
+Asset packages, USD files, materials, textures, and validation reports are available
+in [Files and versions](https://huggingface.co/datasets/nvidia/simready-assets/tree/main).
+Refer to each package's license and metadata for its terms and details.
+
+The thumbnail catalog is generated from the asset packages. See
+[catalog maintenance](catalog/MAINTENANCE.md) for refreshing it after asset changes.

catalog/MAINTENANCE.md

@@ -0,0 +1,29 @@
+# Refresh the asset catalog
+
+The catalog contains one row per package, using its main USD entrypoint and the
+matching preview in `simready_usd/.thumbs/256x256/`. Images are embedded in the
+catalog as previews up to 256 pixels wide and tall. Source files stay in place.
+
+Run the generator after publishing a batch of assets:
+
+```bash
+uv run --with huggingface_hub --with pyarrow --with requests --with Pillow scripts/update_asset_catalog.py --output catalog-build --create-pr
+```
+
+Authenticate first with `hf auth login`, or provide `HF_TOKEN` through the runner's
+secret environment. The command creates a pull request updating the catalog and
+asset count. Merge that pull request to refresh the dataset page.
+
+To preview locally, omit `--create-pr`. The generator checks every package has one
+main USD and a matching readable thumbnail. It stops with an error if a package
+cannot be represented, so assets are never silently omitted.
+
+For automatic refreshes, invoke this command as a final step of the existing
+publishing pipeline, or from a webhook runner. Configure the runner to respond to
+asset-package changes on `main`, skip catalog-only changes, and avoid opening
+duplicate refresh pull requests. A schedule or webhook is not installed by this
+change.
+
+The dataset card's configuration explicitly selects `catalog/assets.parquet` for
+the viewer. It does not move, rename, or rewrite asset packages. Dataset access
+settings continue to apply.

catalog/assets.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ee741743ab3b11d0568f1629111d21b562d569232c2682f7b4ed3f3ed483e9d1
+size 76043244

scripts/update_asset_catalog.py

@@ -0,0 +1,201 @@
+"""Build the Hugging Face thumbnail catalog without modifying asset packages.
+
+Dependencies: huggingface_hub, pyarrow, requests, Pillow.
+Uses the existing Hugging Face login or HF_TOKEN. Never stores credentials.
+Run after publishing assets; --create-pr proposes changed catalog files on the Hub.
+"""
+from __future__ import annotations
+
+import argparse
+from collections import defaultdict
+from concurrent.futures import ThreadPoolExecutor, as_completed
+import hashlib
+import io
+import json
+from pathlib import Path, PurePosixPath
+import threading
+from urllib.parse import quote
+
+from huggingface_hub import HfApi, CommitOperationAdd, get_token
+import pyarrow as pa
+import pyarrow.parquet as pq
+from PIL import Image
+import requests
+from requests.adapters import HTTPAdapter
+from urllib3.util.retry import Retry
+
+REPO = "nvidia/simready-assets"
+LOCAL = threading.local()
+
+
+def session():
+    if not hasattr(LOCAL, "session"):
+        s = requests.Session()
+        token = get_token()
+        if token:
+            s.headers["Authorization"] = "Bearer " + token
+        s.mount("https://", HTTPAdapter(max_retries=Retry(
+            total=3, backoff_factor=1, status_forcelist=[429, 500, 502, 503, 504])))
+        LOCAL.session = s
+    return LOCAL.session
+
+
+def discover(paths):
+    files = set(paths)
+    entrypoints_by_package = defaultdict(list)
+    for path in files:
+        if path.count("/") == 2 and "/simready_usd/" in path and PurePosixPath(path).suffix.lower() in (".usd", ".usda", ".usdc"):
+            entrypoints_by_package[path.split("/")[0]].append(path)
+    packages = sorted(p.rsplit("/", 1)[0] for p in files
+                      if p.count("/") == 1 and p.endswith("/com.nvidia.simready.packaging.json"))
+    rows = []
+    for package in packages:
+        prefix = package + "/simready_usd/"
+        entrypoints = sorted(entrypoints_by_package[package])
+        if len(entrypoints) != 1:
+            raise ValueError(f"Expected one main USD for {package}; found {len(entrypoints)}")
+        asset = entrypoints[0]
+        thumbnail = prefix + ".thumbs/256x256/" + PurePosixPath(asset).name + ".png"
+        if thumbnail not in files:
+            raise ValueError(f"Missing asset thumbnail: {thumbnail}")
+        rows.append({"name": package, "asset_path": asset, "thumbnail_path": thumbnail})
+    if not rows:
+        raise ValueError("No asset packages found")
+    return rows
+
+
+def fetch_thumbnail(row, repo, revision, cache):
+    key = hashlib.sha256((repo + revision + row["thumbnail_path"]).encode()).hexdigest()
+    target = cache / (key + ".png")
+    if target.exists():
+        data = target.read_bytes()
+    else:
+        url = f"https://huggingface.co/datasets/{repo}/resolve/{revision}/{quote(row['thumbnail_path'], safe='/')}"
+        response = session().get(url, timeout=(15, 60))
+        response.raise_for_status()
+        data = response.content
+        target.write_bytes(data)
+    with Image.open(io.BytesIO(data)) as image:
+        image.verify()
+    with Image.open(io.BytesIO(data)) as image:
+        image.thumbnail((256, 256), Image.Resampling.LANCZOS)
+        preview = io.BytesIO()
+        image.save(preview, format="PNG", optimize=True)
+        return preview.getvalue()
+
+
+def card_text(repo, count):
+    return f"""---
+pretty_name: NVIDIA SimReady Assets
+tags:
+- 3d
+- simready
+configs:
+- config_name: catalog
+  default: true
+  data_files:
+  - split: assets
+    path: catalog/assets.parquet
+dataset_info:
+- config_name: catalog
+  features:
+  - name: thumbnail
+    dtype: image
+  - name: name
+    dtype: string
+  - name: asset_url
+    dtype: string
+---
+
+# NVIDIA SimReady Assets
+
+Browse **{count:,} assets** in the thumbnail table on this page. Each row shows the
+asset's existing preview, its package name, and a link to its main USD file.
+
+[Open the full asset table](https://huggingface.co/datasets/{repo}/viewer/catalog/assets)
+to search by name and navigate between pages.
+
+Asset packages, USD files, materials, textures, and validation reports are available
+in [Files and versions](https://huggingface.co/datasets/{repo}/tree/main).
+Refer to each package's license and metadata for its terms and details.
+
+The thumbnail catalog is generated from the asset packages. See
+[catalog maintenance](catalog/MAINTENANCE.md) for refreshing it after asset changes.
+"""
+
+
+def build(args):
+    api = HfApi()
+    info = api.dataset_info(args.repo, revision=args.revision)
+    revision = info.sha
+    rows = discover([s.rfilename for s in info.siblings])
+    output = Path(args.output).resolve()
+    cache = output / ".cache" / "thumbnails"
+    cache.mkdir(parents=True, exist_ok=True)
+    records = [None] * len(rows)
+    print(f"Found {len(rows)} packages at {revision}", flush=True)
+    with ThreadPoolExecutor(max_workers=args.workers) as pool:
+        futures = {pool.submit(fetch_thumbnail, row, args.repo, revision, cache): i
+                   for i, row in enumerate(rows)}
+        for done, future in enumerate(as_completed(futures), 1):
+            i = futures[future]
+            row = rows[i]
+            records[i] = {
+                "thumbnail": {"bytes": future.result(), "path": row["thumbnail_path"]},
+                "name": row["name"],
+                "asset_url": f"https://huggingface.co/datasets/{args.repo}/blob/main/{quote(row['asset_path'], safe='/')}",
+            }
+            if done % 100 == 0 or done == len(rows):
+                print(f"Validated thumbnails: {done}/{len(rows)}", flush=True)
+    schema = pa.schema([
+        ("thumbnail", pa.struct([("bytes", pa.binary()), ("path", pa.string())])),
+        ("name", pa.string()), ("asset_url", pa.string()),
+    ], metadata={b"huggingface": json.dumps({"info": {"features": {
+        "thumbnail": {"_type": "Image"},
+        "name": {"dtype": "string", "_type": "Value"},
+        "asset_url": {"dtype": "string", "_type": "Value"},
+    }}}).encode()})
+    table = pa.Table.from_pylist(records, schema=schema)
+    catalog = output / "catalog" / "assets.parquet"
+    catalog.parent.mkdir(parents=True, exist_ok=True)
+    pq.write_table(table, catalog, compression="zstd", row_group_size=50, write_page_index=True)
+    readback = pq.read_table(catalog)
+    assert readback.num_rows == len(rows)
+    assert readback.column("name").to_pylist() == [r["name"] for r in rows]
+    assert all(r["thumbnail"]["bytes"] for r in readback.to_pylist())
+    # Preserve the card on future refreshes; update only its generated count.
+    readme = output / "README.md"
+    try:
+        response = session().get(f"https://huggingface.co/datasets/{args.repo}/resolve/{revision}/README.md", timeout=30)
+        if response.status_code == 404:
+            text = card_text(args.repo, len(rows))
+        else:
+            response.raise_for_status()
+            import re
+            text = response.text
+            if "path: catalog/assets.parquet" not in text:
+                raise ValueError("Existing README has no catalog configuration; integrate it manually")
+            text = re.sub(r"Browse \*\*[\d,]+ assets\*\*", f"Browse **{len(rows):,} assets**", text)
+        readme.write_text(text, encoding="utf-8")
+    except requests.RequestException:
+        raise
+    audit = {"repo": args.repo, "revision": revision, "rows": len(rows),
+             "parquet_bytes": catalog.stat().st_size, "assets": rows}
+    (output / "audit.json").write_text(json.dumps(audit, indent=2), encoding="utf-8")
+    print(json.dumps({k: v for k, v in audit.items() if k != "assets"}), flush=True)
+    if args.create_pr:
+        result = api.create_commit(repo_id=args.repo, repo_type="dataset", parent_commit=revision,
+            operations=[CommitOperationAdd(path_in_repo="catalog/assets.parquet", path_or_fileobj=str(catalog)),
+                        CommitOperationAdd(path_in_repo="README.md", path_or_fileobj=str(readme))],
+            commit_message="Refresh asset thumbnail catalog", create_pr=True)
+        print(result.pr_url)
+
+
+if __name__ == "__main__":
+    p = argparse.ArgumentParser(description=__doc__)
+    p.add_argument("--repo", default=REPO)
+    p.add_argument("--revision", default="main")
+    p.add_argument("--output", default="catalog-build")
+    p.add_argument("--workers", type=int, default=8)
+    p.add_argument("--create-pr", action="store_true")
+    build(p.parse_args())