facebookresearch/detectron2 · error · KeyError

No built-in metadata for dataset {}

Error message

No built-in metadata for dataset {}

What it means

detectron2's builtin metadata registry only knows a fixed set of dataset names (COCO, cityscapes, etc.). _get_builtin_metadata raises KeyError when asked for metadata of a dataset name that has no builtin entry.

Source

Thrown at detectron2/data/datasets/builtin_meta.py:350

            "keypoint_connection_rules": KEYPOINT_CONNECTION_RULES,
        }
    elif dataset_name == "cityscapes":
        # fmt: off
        CITYSCAPES_THING_CLASSES = [
            "person", "rider", "car", "truck",
            "bus", "train", "motorcycle", "bicycle",
        ]
        CITYSCAPES_STUFF_CLASSES = [
            "road", "sidewalk", "building", "wall", "fence", "pole", "traffic light",
            "traffic sign", "vegetation", "terrain", "sky", "person", "rider", "car",
            "truck", "bus", "train", "motorcycle", "bicycle",
        ]
        # fmt: on
        return {
            "thing_classes": CITYSCAPES_THING_CLASSES,
            "stuff_classes": CITYSCAPES_STUFF_CLASSES,
        }
    raise KeyError("No built-in metadata for dataset {}".format(dataset_name))

View on GitHub (pinned to a2f4a8771a)

Solutions

1. Register your own metadata: MetadataCatalog.set('mydataset_train', thing_classes=[...]) before use
2. Check the exact dataset name spelling against the keys in detectron2/data/datasets/builtin_meta.py
3. For custom COCO-format data use register_coco_instances which sets metadata from the json's categories

Example fix

# before
meta = MetadataCatalog.get('mydata_train')
# after
from detectron2.data import MetadataCatalog
MetadataCatalog.set('mydata_train', thing_classes=['cat', 'dog'])
meta = MetadataCatalog.get('mydata_train')

from detectron2.data import MetadataCatalog
name = 'mydata_train'
if not MetadataCatalog.contains(name):
    MetadataCatalog.set(name, thing_classes=['cat', 'dog'])

try:
    meta = MetadataCatalog.get(name)
except (KeyError, AssertionError):
    MetadataCatalog.set(name, thing_classes=classes)
    meta = MetadataCatalog.get(name)

When it happens

Trigger: Calling MetadataCatalog.get(name) (or a registration helper) for a name like 'mydataset_train' without registering custom metadata first; or a typo/varied name such as 'coco_2017_trains' or 'cityscapes_fine_*_val' variant not covered by builtin_meta.py keys.

Common situations: Training on a custom or renamed dataset without MetadataCatalog.set(); dataset name typos; splitting/renaming builtin datasets (e.g. 'coco_2017_val_100') that are expected to hit builtin entries but miss the key.

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
• Encountered category_id={annotation_category_id} but this id
• Cannot match one checkpoint key to multiple keys in the mode

AI-assisted analysis of facebookresearch/detectron2@a2f4a8771a (2026-08-27). Data as JSON: /api/errors/1598f9ceeeb7acac. Report an issue: GitHub.