Training engine
elfes.engine runs the training and evaluation of PyTorch models, from parameter updates to checkpoint selection and resuming a paused run. A workflow defines how a model acts on a batch and which errors it should learn from. The engine carries out that training procedure on CPU, CUDA, or multiple devices, with consistent loss normalization, metric aggregation, and data progress tracking.
Training and optimization starts with a complete training example and explains update schedules and training state. Objectives and metrics describes how a step reports losses and statistics, including physical quantities that need different averaging weights.
Evaluation covers validation, averaged parameters, and best-model selection. Checkpoints and resuming follows a run through stopping, continuation, and loading a saved model. Run logging describes the configuration, metrics, and events recorded along the way.