Loss Function¶

The detection loss lives in yolo/tasks/detection/loss.py. It is assembled from three composable components and wrapped in a dual-loss strategy for auxiliary and main heads.

Components¶

Class	Description
`BCELoss`	Binary cross-entropy on class predictions, normalized by a per-batch class factor
`BoxLoss`	IoU-based box regression loss (CIoU by default)
`DFLoss`	Distribution Focal Loss on the regression distribution predicted by DFL heads
`YOLOLoss`	Combines `BCELoss` + `BoxLoss` + `DFLoss` for a single prediction head. Handles anchor separation and target assignment internally
`DualLoss`	Runs `YOLOLoss` for both the auxiliary and main prediction heads, weighting the auxiliary loss at 0.25×

The entry point used by the solver is:

from yolo.tasks.detection.loss import create_loss_function

loss_fn = create_loss_function(cfg, vec2box)

create_loss_function builds a DualLoss configured from cfg.task.loss (LossConfig).

Loss Config¶

Loss weights are controlled via yolo/config/task/train.yaml under the loss key:

loss:
  box: 7.5
  cls: 0.5
  dfl: 1.5

Override from the CLI:

python -m yolo task=train task.loss.box=10.0

Target Assignment¶

YOLOLoss uses BoxMatcher (from yolo/tasks/detection/postprocess.py) to assign ground-truth boxes to anchors via a Task-Aligned Assigner (TAL). Loss is computed only on matched anchors.