loss ==== The ``loss`` config group determines all loss functions that are used for ``mml`` models. It determines a single loss function per task type. More precisely the :class:`~mml.core.data_loading.task_attributes.TaskType` is matched to a short string via :attr:`~mml.core.models.lightning_single_frame.CONFIG_ROUTES`. For each model head (that matches a task type) the respective `loss.SHORT_STRING` config is instantiated. The main file determines some additional behaviour: default ~~~~~~~ The default top level configuration. .. autoyaml:: loss/default.yaml The following task type matching subdirectories exist: cls --- For classification tasks, currently shipped with a single option ce ~~ .. autoyaml:: loss/cls/ce.yaml mlcls ----- For multi-label classification tasks, currently shipped with two options bce ~~~ .. autoyaml:: loss/mlcls/bce.yaml ce ~~ .. autoyaml:: loss/mlcls/ce.yaml reg --- For regression tasks, currently shipped with a single option huber ~~~~~ .. autoyaml:: loss/reg/huber.yaml seg --- For segmentation tasks, currently shipped with two options ce ~~ .. autoyaml:: loss/seg/ce.yaml dice ~~~~ .. autoyaml:: loss/seg/dice.yaml