metrics

The metrics config group determines all metrics that are used to evaluate mml models. It determines a suite of metrics per task type. More precisely the TaskType is matched to a short string via CONFIG_ROUTES. For each model head (that matches a task type) the respective metrics.SHORT_STRING config is instantiated.

The following task type matching subdirectories exist and contain a single default.yaml file each:

  • cls (for classification tasks)

  • mlcls (for multi-label classification tasks)

  • reg (for regression tasks)

  • seg (for segmentation tasks)

The currently used metrics can be seen via mml --cfg=job (add `` | grep torchmetrics`` for reduced output). All metrics used by mml are from torchmetrics.

The main file determines some additional behaviour:

default

The default top level configuration.

bootstrap
default: 20