opendatalab/MinerU · error · ValueError

PPDocLayoutV2ForObjectDetection only supports inference.

Error message

PPDocLayoutV2ForObjectDetection only supports inference.

What it means

ValueError raised in PPDocLayoutV2ForObjectDetection.forward when labels is not None. This DETR-style detection head is shipped for inference only — the training/loss branch is deliberately absent — so supplying labels (the HF training convention) is rejected with a clear message.

Source

Thrown at mineru/model/layout/pp_doclayoutv2.py:828

            persistent=False,
        )
        self.post_init()

    def forward(
        self,
        pixel_values: torch.FloatTensor,
        pixel_mask: Optional[torch.LongTensor] = None,
        encoder_outputs: Optional[torch.FloatTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        decoder_inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[list[dict]] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        **kwargs,
    ):
        if labels is not None:
            raise ValueError("PPDocLayoutV2ForObjectDetection only supports inference.")

        use_return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        outputs = self.model(
            pixel_values,
            pixel_mask=pixel_mask,
            encoder_outputs=encoder_outputs,
            inputs_embeds=inputs_embeds,
            decoder_inputs_embeds=decoder_inputs_embeds,
            labels=None,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=True,
        )

        raw_bboxes = outputs.intermediate_reference_points[:, -1]
        logits = outputs.intermediate_logits[:, -1]

        box_centers, box_sizes = raw_bboxes.split(2, dim=-1)

View on GitHub (pinned to 4fe4bde114)

Solutions

1. Call forward without labels: model(pixel_values=..., pixel_mask=...).
2. For evaluation, compute metrics from the returned detections instead of model loss.
3. For fine-tuning, use the original upstream PP-DocLayout training codebase rather than this inference-only wrapper.

Example fix

# before
outputs = model(pixel_values=pv, pixel_mask=pm, labels=targets)  # ValueError

# after
outputs = model(pixel_values=pv, pixel_mask=pm)  # inference only

def inference_call(model, pixel_values, pixel_mask=None, **kwargs):
    kwargs.pop('labels', None)  # training-style args are unsupported
    return model(pixel_values=pixel_values, pixel_mask=pixel_mask, **kwargs)

try:
    outputs = model(pixel_values=pv, pixel_mask=pm, labels=targets)
except ValueError as e:
    if 'only supports inference' in str(e):
        outputs = model(pixel_values=pv, pixel_mask=pm)  # retry without labels
    else:
        raise

When it happens

Trigger: model(pixel_values, labels=[{'class_labels': ..., 'boxes': ...}]) — i.e. calling forward the way you would with transformers' DetrForObjectDetection during training or eval-with-loss.

Common situations: Reusing HF detection training loops against this model; evaluation scripts that compute loss from labels; fine-tuning attempts on a checkpoint that only supports inference.

Related errors

• PP-DocLayoutV2 reading-order mask must be 2D or 4D, got shap
• The hidden size ({config.hidden_size}) is not a multiple of
• PP-DocLayoutV2 reading-order inference requires a mask tenso
• Attention mask batch size {attention_mask.shape[0]} does not
• Unsupported image type for PP-DocLayoutV2: {type(image)}

AI-assisted analysis of opendatalab/MinerU@4fe4bde114 (2026-08-14). Data as JSON: /api/errors/18a7c8fe00257488. Report an issue: GitHub.