What causes "[DBLoss]: Unrecognized main loss type!" in PaddlePaddle/PaddleOCR?

Set main_loss_type to 'BCELoss' or 'DiceLoss' in the DBLoss config

PaddlePaddle/PaddleOCR · error · Exception

[DBLoss]: Unrecognized main loss type!

Error message

[DBLoss]: Unrecognized main loss type!

What it means

DBLoss.__init__ branches on main_loss_type and only wires up 'BCELoss' (bce_loss path with BalanceLoss) and 'DiceLoss' (BalanceLoss + DiceLoss). Even though BalanceLoss itself supports five types, DBLoss as a whole only accepts these two, and anything else raises Exception('[DBLoss]: Unrecognized main loss type!').

Source

Thrown at ppocr/losses/det_db_loss.py:74

        self.l1_loss = MaskL1Loss(eps=eps)
        if main_loss_type == "DiceFocalLoss":
            self.bce_loss = DiceFocalLoss(
                dice_weight=dice_weight,
                focal_weight=focal_weight,
                focal_alpha=focal_alpha,
                focal_gamma=focal_gamma,
                eps=eps,
            )
            self.dice_loss = self.bce_loss
        elif main_loss_type == "DiceLoss":
            self.bce_loss = BalanceLoss(
                balance_loss=balance_loss,
                main_loss_type=main_loss_type,
                negative_ratio=ohem_ratio,
            )
            self.dice_loss = DiceLoss(eps=eps)
        else:
            raise Exception("[DBLoss]: Unrecognized main loss type!")
        self.aux_weight_p4 = aux_weight_p4
        self.aux_weight_p3 = aux_weight_p3
        self.aux_weight_p2 = aux_weight_p2

    def forward(self, predicts, labels):
        predict_maps = predicts["maps"]
        (
            label_threshold_map,
            label_threshold_mask,
            label_shrink_map,
            label_shrink_mask,
        ) = labels[1:]
        shrink_maps = predict_maps[:, 0, :, :]
        threshold_maps = predict_maps[:, 1, :, :]
        binary_maps = predict_maps[:, 2, :, :]

        loss_shrink_maps = self.bce_loss(
            shrink_maps, label_shrink_map, label_shrink_mask

View on GitHub (pinned to 2661c7c0ef)

Solutions

Set main_loss_type to 'BCELoss' or 'DiceLoss' in the DBLoss config
Verify you are editing the Loss section of the actual det model config being loaded (not a base config that is overridden)

Example fix

# before (config yaml)
Loss:
  name: DBLoss
  main_loss_type: MaskL1Loss

# after
Loss:
  name: DBLoss
  main_loss_type: DiceLoss

Defensive patterns

Strategy: validation

Validate before calling

DB_LOSS_TYPES = {'BCELoss', 'DiceLoss'}
assert main_loss_type in DB_LOSS_TYPES, f'DBLoss supports only {DB_LOSS_TYPES}, got {main_loss_type!r}'

Type guard

def is_db_loss_type(t: str) -> bool:
    return t in {'BCELoss', 'DiceLoss'}

Prevention

Note the narrower set vs BalanceLoss: DBLoss accepts only BCELoss/DiceLoss
Assert the loss config the moment it is parsed so the failure points at the YAML, not at model build
Run a 1-iteration smoke train on config changes

When it happens

Trigger: Constructing DBLoss(main_loss_type='MaskL1Loss') or 'CrossEntropy' — values legal for BalanceLoss but not for DBLoss; typically via a det_db config YAML.

Common situations: Copying a BalanceLoss-supported type into a DB config, or upgrading configs from another detection model where those types were valid, then hitting a confusing error because the inner class accepts them.

Related errors

AI-assisted analysis of PaddlePaddle/PaddleOCR@2661c7c0ef (2026-08-14). Data as JSON: /api/errors/ec178980cf19e6f4. Report an issue: GitHub.