engine¶

The xuance.engine package provides the backend-agnostic execution layer of Xuance. It is responsible for orchestrating the lifecycle of an experiment, including training, evaluation, benchmarking, logging, and result management, while remaining independent of the underlying deep learning framework.

In Xuance, the Engine coordinates how an experiment runs, whereas backend implementations (xuance.torch, xuance.tensorflow, xuance.mindspore) define how algorithms are implemented. This separation allows the same experiment workflow to be executed across multiple backends without modifying the high-level logic.

Key responsibilities of the Engine include:

Managing experiment modes such as train, test, and benchmark.
Creating environments and connecting them with backend-specific agents.
Driving the interaction loop by invoking agent.train() or agent.test().
Handling logging, checkpointing, and result persistence.
Providing task-specific runners for different environments (e.g., DRL, multi-agent settings, or domain-specific benchmarks).

By isolating orchestration from implementation, Xuance ensures a clean separation between experiment control and algorithm realization, improving maintainability, extensibility, and cross-backend reproducibility.