mml.core.models

MML abstracts it’s models into underlying torch models and overlaying lightning models. The torch model encapsulates the network description while the lightning wrapper controls loss computation, optimizer, metrics and logging.

This allows very simple extensions via adding new torch models and wrapping those with the same lightning module. Two examples for very flexible torch models are the timm and smp models that build upon the timm library and the pytorch_segmentation_models library respectively. So far only single frame tasks are supported by the lightning_single_frame lightning wrapper but video clip support is planned.