FastUnifiedInstanceLoss

class FastUnifiedInstanceLoss(cost_mask: float = 1.0, cost_dice: float = 1.0, cost_class: float = 0.0, num_points: int = 0, ignore_index: int = -1, loss_weight_focal: float = 1.0, loss_weight_dice: float = 1.0, cls_weight_matched: float = 2.0, cls_weight_noobj: float = 0.1, focal_alpha: float = 0.25, focal_gamma: float = 2.0, truth_label: str = 'instance', segment_key: str = 'segment', aux_loss_weight: float = 1.0, *, query_heads, overlap: bool = False, membership_key: str = 'inst_pe', valid_key: str = 'ring_valid', mask_pe_thresh: float = 0.0, noobj_mask_loss_weight: float = 0.0)

Bases: Module

Combine instance mask/class losses with configured query-head losses.

Non-overlap mode builds masks from per-point instance ids and aggregates each matched instance’s point targets for its query heads. Overlap mode uses the configured multi-hot membership matrix, validity mask, and per-instance head targets. aux_outputs provide deep supervision and are scaled by aux_loss_weight.

query_heads is a list of named mappings. Continuous heads use a regression criterion; categorical heads use cross-entropy. Each mapping selects its prediction key, target key, target aggregation, and loss weight.

FastUnifiedInstanceLoss.forward(pred: Dict, input_dict: Dict)