Configuration

This page describes how to configure the metrics module used to score model predictions.

Options

Name	Allowed	Default	Description
_target_	"ClassificationMetrics", "DetectionMetrics"	"ClassificationMetrics"	The type of metrics to use. See Underlying libraries for more details.
task	"binary", "multiclass", "multilabel"	"multiclass"	Classification task type. Only used by ClassificationMetrics.
num_classes	integer, null	null	Number of classes for multiclass tasks. For multilabel, this is also accepted as an alias for num_labels.
num_labels	integer, null	null	Number of labels for multilabel tasks.
average	"micro", "macro", "weighted", "none"	"macro"	Aggregation mode. ClassificationMetrics supports "micro", "macro", "weighted", and "none" for multiclass and multilabel. DetectionMetrics supports "micro" and "macro". See Underlying libraries for more details.
ignore_index	integer, null	null	Optional target value to ignore when computing classification metrics.
threshold	float, null	0.5 for multilabel when omitted	Threshold applied by ClassificationMetrics for multilabel predictions. Ignored by other targets unless the underlying TorchMetrics implementation accepts it through forwarded kwargs.
box_format	"xyxy", "xywh"	"xyxy"	Bounding-box format expected by DetectionMetrics.
iou_type	"bbox", "segm", or a tuple containing those values	"bbox"	IoU mode passed to detection mean average precision.
iou_thresholds	list[float], null	null	Optional custom IoU thresholds for detection evaluation.
rec_thresholds	list[float], null	null	Optional custom recall thresholds for detection evaluation.
max_detection_thresholds	list[int], null	null	Optional maximum-detection cutoffs for detection evaluation.
class_metrics	true, false	false	Whether DetectionMetrics should compute class-wise metrics in addition to global summaries.
extended_summary	true, false	false	Whether DetectionMetrics should request extended non-scalar summary outputs.
backend	"faster_coco_eval", "pycocotools"	"faster_coco_eval"	Backend used by detection mean average precision. pycocotools should be used only for backward compatibility.

YAML example

metrics:
  _target_: "ClassificationMetrics"
  task: "multiclass"
  num_classes: 7
  num_labels: null
  average: "macro"
  ignore_index: null

CLI override example

uv

uv run raitap metrics=detection metrics.class_metrics=true metrics.extended_summary=true

pip

raitap metrics=detection metrics.class_metrics=true metrics.extended_summary=true