Losses

from nobg.loss import birefnet_loss, sam3_loss

Every loss has the same signature — a list of prediction tensors and one ground truth — so it can be assigned to model.criterion as-is:

def my_loss(scaled_preds: list[Tensor], gt: Tensor) -> Tensor: ...

model.criterion = my_loss

scaled_preds are raw logits, not probabilities: BiRefNet passes its final scale plus every intermediate decoder scale, Sam3 always passes a single-element list. Any prediction whose spatial size differs from gt is bilinearly resized to match, so the intermediate scales need no handling from the caller. gt is (B, 1, H, W) in [0, 1].

criterion is a plain function attribute, not a submodule — it never enters the state dict, and replacing it is just an assignment.

birefnet_loss

birefnet_loss(scaled_preds, gt) -> Tensor

The default for BiRefNet, matching the paper's training recipe. Each scale contributes three terms:

Summed over every scale — so the value grows with the number of decoder scales; it is a training signal, not a comparable metric.

iou_loss

iou_loss(pred, target) -> Tensor

1 - soft IoU, per-sample over spatial dims then batch-averaged. Expects probabilities.

ssim_loss

ssim_loss(pred, target) -> Tensor

(1 - SSIM) / 2 clamped to [0, 1], over 3×3 windows via avg_pool2d, with the standard C1 = 0.01², C2 = 0.03². Expects probabilities. This is the term that pushes local structure — edges and thin detail — rather than per-pixel correctness.

sam3_loss

sam3_loss(scaled_preds, gt, *, focal_alpha=0.6, focal_gamma=2.0,
          focal_weight=20.0, dice_weight=30.0) -> Tensor

The default for Sam3 — SAM 3's semantic-segmentation objective, so fine-tuning matches how the checkpoint was trained: 20 × focal + 30 × dice at Meta's published relative weights.

The keyword arguments make it tunable without rewriting it:

from functools import partial

model.criterion = partial(sam3_loss, dice_weight=10.0)

sigmoid_focal_loss

sigmoid_focal_loss(pred, target, alpha=0.6, gamma=2.0) -> Tensor

Focal loss (arXiv:1708.02002) on raw logits, mean-reduced. Note alpha=0.6, not RetinaNet's 0.25: SAM 3's semantic head favours the foreground, because a matte's positive class covers a large fraction of the image rather than a handful of anchors. Pass alpha=-1 to disable the class weighting entirely.

dice_loss

dice_loss(pred, target) -> Tensor

Soft dice on raw logits: 1 - 2|X∩Y| / (|X|+|Y|), smoothed by 1, per-sample over flattened spatial dims then batch-averaged. The mask-overlap term of DETR-family set losses, which SAM 3 inherits.

Choosing

Mixing is legal — BiRefNet.criterion = sam3_loss works, since the signature is shared — but you are then no longer matching the recipe the weights were trained under.

Evaluating, as opposed to training, is Metrics.