sop-e2e-inference

Run end-to-end SOP model evaluations by submitting temporal segmentation and action recognition jobs to the evaluation-ms HTTP API. Configure inference parameters, model checkpoints, and chunking strategies via a YAML file to validate trained vision-language models.

Submitting SOP monitoring evaluation jobs requires a reachable evaluation-ms service, valid training and dataset job identifiers, PyYAML for configuration parsing, and host-side access to retrieve generated evaluation artifacts like accuracy and log files.

Yes, SOP benchmarking supports both DDM-based temporal segmentation and uniform video chunking. You can control model checkpoints, inference parameters, thresholds, resolution, GPU selection, and timeouts through configurable pipeline settings.

Monitor asynchronous evaluation jobs through the evaluation-ms HTTP API until completion. The process resolves host-side output directories, surfaces headline metrics like temporal-segmentation F1 and action-recognition accuracy, and guides you to detailed per-video results.

Yes, SOP end-to-end inference allows you to specify GPU selection and resolution settings. These configurable inference parameters are applied during DDM or uniform video chunking to ensure accurate action recognition and temporal segmentation evaluation.

The evaluation-ms HTTP API requires a reachable service instance and valid dataset job identifiers. Without host-side access to output directories, users cannot retrieve detailed segmentation, accuracy, and log artifacts generated during the asynchronous evaluation.