PaddlePaddle/PaddleOCR · error · ValueError

error!!!!!!

Error message

error!!!!!!

What it means

This is the fallback branch of a distillation loss's forward: it dispatches on self.mode, which must be one of 'log', 'mean', 'sum', 'meanlog' or 'ctcdkd'. An unrecognized mode reaches the else and raises the unhelpful ValueError('error!!!!!!') — the message gives no clue, but the cause is always an invalid mode string on this specific loss class (the mode selection in the distillation YAML, e.g. for the ctc/attention distill loss).

Source

Thrown at ppocr/losses/distillation_loss.py:1139

        elif self.mode == "sum":
            return self.forward_sum(stu_out, tea_out)
        elif self.mode == "meanlog":
            blank_mask = paddle.ones_like(stu_out)
            blank_mask.stop_gradient = True
            blank_mask[:, :, 0] = -1
            stu_out *= blank_mask
            tea_out *= blank_mask
            return self.forward_meanlog(stu_out, tea_out)
        elif self.mode == "ctcdkd":
            # ignore ctc blank logits
            blank_mask = paddle.ones_like(stu_out)
            blank_mask.stop_gradient = True
            blank_mask[:, :, 0] = -1
            stu_out *= blank_mask
            tea_out *= blank_mask
            return self.ctc_dkd_loss(stu_out, tea_out, targets)
        else:
            raise ValueError("error!!!!!!")

    def forward_log(self, out1, out2):
        if self.act is not None:
            out1 = self.act(out1) + 1e-10
            out2 = self.act(out2) + 1e-10
        if self.use_log is True:
            # for recognition distillation, log is needed for feature map
            log_out1 = paddle.log(out1)
            log_out2 = paddle.log(out2)
            loss = (self._kldiv(log_out1, out2) + self._kldiv(log_out2, out1)) / 2.0

        return loss


class DistillCTCLogits(KLCTCLogits):
    def __init__(
        self, model_name_pairs=[], key=None, name="ctc_logits", reduction="mean"
    ):

View on GitHub (pinned to 2661c7c0ef)

Solutions

1. Find the distillation loss config whose forward hits this line and set mode to one of: 'log', 'mean', 'sum', 'meanlog', 'ctcdkd'
2. Match the mode to the loss class: 'ctcdkd' belongs to the CTC distill loss; check the loss's own mode branches before choosing
3. Add a quick assert at startup: assert loss.mode in {'log','mean','sum','meanlog','ctcdkd'} to fail at config time instead of mid-training

Example fix

# before (distillation config yaml)
mode: kl        # ValueError: 'error!!!!!!' at first step

# after
mode: meanlog

DISTILL_MODES = {'log', 'mean', 'sum', 'meanlog', 'ctcdkd'}
assert cfg['mode'] in DISTILL_MODES, f"distill loss mode must be one of {sorted(DISTILL_MODES)}, got {cfg['mode']!r}"

def is_distill_mode(m: str) -> bool:
    return isinstance(m, str) and m in {'log', 'mean', 'sum', 'meanlog', 'ctcdkd'}

try:
    loss = distill_loss(student_out, teacher_out, targets)
except ValueError as e:
    if 'error' in str(e):
        raise ValueError(f'Invalid distillation mode {distill_loss.mode!r}; expected one of log/mean/sum/meanlog/ctcdkd') from e
    raise

When it happens

Trigger: Building a DistillationModel loss item whose mode is misspelled or from a different loss class (e.g. mode: ' CEL ' with spaces, or a mode valid for KLJSLoss like 'kl' but not for this loss), then running forward during training.

Common situations: Hand-editing a distillation config and reusing a mode from another distill loss; mode/config mismatch after upgrading PaddleOCR where supported mode names changed; the error surfaces only at the first training step, not at config parse time.

Related errors

• The mode.lower() if KLJSLoss should be one of ['kl', 'js']
• main_loss_type in BalanceLoss() can only be one of {}
• [DBLoss]: Unrecognized main loss type!
• {} is not supported in MultiLoss yet
• RecResizeImg.image_shape is required in rec inference.yml

AI-assisted analysis of PaddlePaddle/PaddleOCR@2661c7c0ef (2026-08-14). Data as JSON: /api/errors/1cf963bb82a0cd16. Report an issue: GitHub.