---

name: datasets-evaluation description: "Prepare and validate MMDetection datasets, annotation formats, dataloaders, transforms, samplers, evaluators, and analysis workflows." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Datasets Evaluation

Use this sub-skill when a task is about dataset layout, annotation conversion, custom dataset config, transform pipelines, sampler choices, evaluator metrics, dataset browsing, or post-test analysis.

Route First

• Dataset files or annotations: use references/data-formats.md for COCO, panoptic, MMEngine middle format, data_root, ann_file, data_prefix, metainfo, class order, transforms, samplers, and tiny validation checks.
• Metrics or analysis: use references/evaluation-analysis.md for CocoMetric, VOCMetric, CityScapesMetric, evaluator config, format_only, classwise AP, browsing, error analysis, and robustness commands.
• Failures: use references/troubleshooting.md for schema, path, category, palette, transform-key, dependency, and metric-output diagnosis.
• Image folder manifest: use scripts/images_to_coco.py to create a COCO-like JSON with image and category entries but no bounding-box annotations.

Boundaries

• For launching training, testing, distributed jobs, resume, or checkpoint result dumping, route to training-testing.
• For custom dataset class implementation or registration code, route to customization-extension after this sub-skill defines the target format and config contract.
• For inference visualization and prediction rendering, route to inference-visualization.
• For selecting or editing base model configs outside dataset/evaluator fields, route to configuration-model-zoo.

Minimal Custom COCO Config Pattern

dataset_type = 'CocoDataset'
data_root = 'data/my_dataset/'
metainfo = dict(classes=('cat', 'dog'), palette=[(220, 20, 60), (0, 0, 142)])

train_dataloader = dict(dataset=dict(
    type=dataset_type,
    data_root=data_root,
    metainfo=metainfo,
    ann_file='annotations/train.json',
    data_prefix=dict(img='train/'),
    filter_cfg=dict(filter_empty_gt=True, min_size=32)))
val_dataloader = dict(dataset=dict(
    type=dataset_type,
    data_root=data_root,
    metainfo=metainfo,
    ann_file='annotations/val.json',
    data_prefix=dict(img='val/'),
    test_mode=True))
test_dataloader = val_dataloader
val_evaluator = dict(type='CocoMetric', ann_file=data_root + 'annotations/val.json', metric='bbox')
test_evaluator = val_evaluator

If the class count changes, update dataset metainfo.classes, every model head num_classes, evaluator ann_file, and visualization palette together.