facebookresearch/detectron2 · error · ValueError

No built-in metadata for dataset {}

Error message

No built-in metadata for dataset {}

What it means

get_lvis_instances_meta resolves metadata by substring matching on the dataset name: names containing 'cocofied' get COCO metadata, 'v0.5' and 'v1' get LVIS versions. Any other LVIS-like name raises ValueError because no builtin metadata matches.

Source

Thrown at detectron2/data/datasets/lvis.py:184

def get_lvis_instances_meta(dataset_name):
    """
    Load LVIS metadata.

    Args:
        dataset_name (str): LVIS dataset name without the split name (e.g., "lvis_v0.5").

    Returns:
        dict: LVIS metadata with keys: thing_classes
    """
    if "cocofied" in dataset_name:
        return _get_coco_instances_meta()
    if "v0.5" in dataset_name:
        return _get_lvis_instances_meta_v0_5()
    elif "v1" in dataset_name:
        return _get_lvis_instances_meta_v1()
    raise ValueError("No built-in metadata for dataset {}".format(dataset_name))


def _get_lvis_instances_meta_v0_5():
    assert len(LVIS_V0_5_CATEGORIES) == 1230
    cat_ids = [k["id"] for k in LVIS_V0_5_CATEGORIES]
    assert min(cat_ids) == 1 and max(cat_ids) == len(
        cat_ids
    ), "Category ids are not in [1, #categories], as expected"
    # Ensure that the category list is sorted by id
    lvis_categories = sorted(LVIS_V0_5_CATEGORIES, key=lambda x: x["id"])
    thing_classes = [k["synonyms"][0] for k in lvis_categories]
    meta = {"thing_classes": thing_classes}
    return meta


def _get_lvis_instances_meta_v1():
    assert len(LVIS_V1_CATEGORIES) == 1203
    cat_ids = [k["id"] for k in LVIS_V1_CATEGORIES]

View on GitHub (pinned to a2f4a8771a)

Solutions

1. Include a recognized substring in the name ('v0.5' or 'v1'), or better, register metadata explicitly
2. For custom LVIS-format data call register_lvis_instances and set MetadataCatalog with your thing_classes/dataset_names
3. Upgrade detectron2 if you need a newer LVIS split

Example fix

# before
meta = get_lvis_instances_meta('lvis_val')
# after
MetadataCatalog.set('lvis_val', thing_classes=my_classes, dataset_name='lvis-val')
meta = MetadataCatalog.get('lvis_val')

from detectron2.data import MetadataCatalog
if not MetadataCatalog.contains(dataset_name):
    raise ValueError(f'register metadata for {dataset_name} before use')

def lvis_meta_available(name: str) -> bool:
    return any(s in name for s in ('cocofied', 'v0.5', 'v1'))

try:
    meta = get_lvis_instances_meta(name)
except ValueError:
    MetadataCatalog.set(name, thing_classes=my_classes, dataset_name='lvis-val')
    meta = MetadataCatalog.get(name)

When it happens

Trigger: Calling get_lvis_instances_meta('lvis_val') (no version substring) or a renamed LVIS dataset like 'lvis_v2_val'; also any call chain through register_all_lvis or load_lvis_json with an unrecognized name.

Common situations: Registering a custom LVIS variant with register_lvis_instances without setting MetadataCatalog yourself; dataset-name typos; newer LVIS versions unsupported by the installed detectron2.

Related errors

• No built-in metadata for dataset {}
• Attribute '{}' does not exist in the metadata of dataset '{}
• Attribute '{key}' does not exist in the metadata of dataset
• Cannot match one checkpoint key to multiple keys in the mode
• Class with @configurable must have a 'from_config' classmeth

AI-assisted analysis of facebookresearch/detectron2@a2f4a8771a (2026-08-27). Data as JSON: /api/errors/44483ab81957e535. Report an issue: GitHub.