First Workflow

This template uses the public classification wrapper. OpenSportsLib does not ship media or a ready-to-run training dataset: supply valid video files and OSL JSON manifests before training.

1. Select a config and prepare splits

Start with opensportslib/configs/classification/video.yaml, or copy it outside the installed package for an experiment. Point DATA.common.splits.<split> at your manifests and media roots.

Create train.json, valid.json, and test.json following the OSL JSON format. A classification sample needs an id, a video input, and labels.action.label; its label must occur in the root label list.

{"version":"2.0","labels":{"action":{"type":"single_label","labels":["pass","shot"]}},"data":[{"id":"clip-001","inputs":[{"type":"video","path":"clips/clip-001.mp4","fps":25.0}],"labels":{"action":{"label":"pass"}}}]}

2. Train, predict, and evaluate

from opensportslib.apis import ClassificationModel

model = ClassificationModel(config="/path/to/classification.yaml")
best_checkpoint = model.train(train_set="/path/to/train.json", valid_set="/path/to/valid.json")
predictions = model.infer(test_set="/path/to/test.json", weights=best_checkpoint)
metrics = model.evaluate(test_set="/path/to/test.json", predictions=predictions)
saved_path = model.save_predictions(output_path="/path/to/predictions.json", predictions=predictions)
print(metrics, saved_path)

infer() returns an in-memory OSL JSON-style payload. save_predictions() is the explicit disk-write step. Use evaluate(..., predictions=saved_path) to score a saved payload without inference.

3. Continue with a task

Use LocalizationModel with a localization config for spotting, or choose one VQA dependency profile before following the VQA guide. See workflows for task-specific arguments and configuration for all config keys.