Configuration¶

The metrics block scores model predictions. The _target_ field is the discriminator: it selects one of four adapters, each with its own set of valid keys.

BinaryClassificationMetrics — two-class problems (accuracy, precision, recall, F1).
MulticlassClassificationMetrics — single-label multiclass classification (one correct class per sample).
MultilabelClassificationMetrics — multilabel classification (multiple correct classes per sample, independent per-label decisions).
DetectionMetrics — object detection (mean average precision and related summaries via torchmetrics).

The previous unified ClassificationMetrics target with a task: binary | multiclass | multilabel field has been removed. Pick the task-specific adapter directly via _target_; each adapter only accepts the keys documented for its section below.

See Underlying libraries for the backend behaviour behind each adapter.

Binary classification¶

Configures BinaryClassificationMetrics for two-class problems.

Options

Name	Allowed	Default	Description
`_target_`	`"BinaryClassificationMetrics"`	`"BinaryClassificationMetrics"`	Selects the binary-classification adapter.
`ignore_index`	`integer, null`	`null`	Optional target value to ignore when computing metrics.
`threshold`	`float`	`0.5`	Decision threshold applied to predicted probabilities to obtain the positive class.

Examples

YAML

metrics:
  _target_: "BinaryClassificationMetrics"

Python

from raitap.metrics import binary_classification

metrics = binary_classification()

CLI

uv run raitap +metrics=binary_classification +metrics.threshold=0.6

pip

raitap +metrics=binary_classification +metrics.threshold=0.6

Multiclass classification¶

Configures MulticlassClassificationMetrics for single-label multiclass problems.

Options

Name	Allowed	Default	Description
`_target_`	`"MulticlassClassificationMetrics"`	`"MulticlassClassificationMetrics"`	Selects the multiclass-classification adapter.
`num_classes`	`integer`	`required`	Number of classes. Must be positive. Required.
`average`	`"micro", "macro", "weighted", "none"`	`"macro"`	Aggregation mode passed to the underlying TorchMetrics implementations. See Underlying libraries for semantics.
`ignore_index`	`integer, null`	`null`	Optional target value to ignore when computing metrics.

Examples

YAML

metrics:
  _target_: "MulticlassClassificationMetrics"
  num_classes: 7

Python

from raitap.metrics import multiclass_classification

metrics = multiclass_classification(num_classes=7)

CLI

uv run raitap +metrics=multiclass_classification +metrics.num_classes=7

pip

raitap +metrics=multiclass_classification +metrics.num_classes=7

Multilabel classification¶

Configures MultilabelClassificationMetrics for multilabel problems (independent per-label decisions).

Options

Name	Allowed	Default	Description
`_target_`	`"MultilabelClassificationMetrics"`	`"MultilabelClassificationMetrics"`	Selects the multilabel-classification adapter.
`num_labels`	`integer`	`required`	Number of labels. Must be positive. Required.
`average`	`"micro", "macro", "weighted", "none"`	`"macro"`	Aggregation mode passed to the underlying TorchMetrics implementations.
`ignore_index`	`integer, null`	`null`	Optional target value to ignore when computing metrics.
`threshold`	`float`	`0.5`	Per-label decision threshold applied to predicted probabilities.

Examples

YAML

metrics:
  _target_: "MultilabelClassificationMetrics"
  num_labels: 5

Python

from raitap.metrics import multilabel_classification

metrics = multilabel_classification(num_labels=5)

CLI

uv run raitap +metrics=multilabel_classification +metrics.num_labels=5

pip

raitap +metrics=multilabel_classification +metrics.num_labels=5

Object detection¶

Configures DetectionMetrics, which wraps torchmetrics MeanAveragePrecision. The IoU-related knobs are grouped under the nested iou: block.

Options

Name	Allowed	Default	Description
`_target_`	`"DetectionMetrics"`	`"DetectionMetrics"`	Selects the object-detection adapter.
`box_format`	`"xyxy", "xywh"`	`"xyxy"`	Bounding-box format of incoming predictions and targets. torchvision detectors output `xyxy`.
`iou.type`	`"bbox", "segm", or a tuple of those values`	`"bbox"`	IoU mode passed to mean average precision.
`iou.thresholds`	`list[float], null`	`null`	Optional custom IoU thresholds. `null` uses the torchmetrics default (COCO-style sweep).
`iou.rec_thresholds`	`list[float], null`	`null`	Optional custom recall thresholds. `null` uses the torchmetrics default.
`iou.max_detection_thresholds`	`list[int], null`	`null`	Optional maximum-detection cutoffs. `null` uses the torchmetrics default.
`class_metrics`	`true, false`	`false`	Whether to compute per-class metrics in addition to global summaries.
`extended_summary`	`true, false`	`false`	Whether to request the torchmetrics extended (non-scalar) summary outputs as artifacts.
`average`	`"macro", "micro"`	`"macro"`	Aggregation mode for detection summaries.
`backend`	`"faster_coco_eval", "pycocotools"`	`"faster_coco_eval"`	Backend used by mean average precision. `pycocotools` should be used only for backward compatibility.

Examples

YAML

metrics:
  _target_: "DetectionMetrics"
  iou:
    thresholds: [0.5, 0.75]
  class_metrics: true

Python

from raitap.metrics import detection

metrics = detection(
    iou={"thresholds": [0.5, 0.75]},
    class_metrics=True,
)

CLI

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