DeprecationWarning: onnx_delta2bbox is deprecated in the cas

Error message

DeprecationWarning: onnx_delta2bbox is deprecated in the case of batch decoding and non-ONNX, please use “delta2bbox” instead. In order to improve the decoding speed, the batch function will no longer be supported.

What it means

Deprecation warning from DeltaXYWHBBoxCoder.decode: the else-branch uses the ONNX export helper onnx_delta2bbox even outside ONNX export when pred_bboxes is a 3-D (batched) tensor. Batched decoding via this path is deprecated for speed; use the vectorized delta2bbox instead, as batch support in onnx_delta2bbox will be dropped.

Source

Thrown at mmdet/models/task_modules/coders/delta_xywh_bbox_coder.py:112

                width and height.

        Returns:
            Union[torch.Tensor, :obj:`BaseBoxes`]: Decoded boxes.
        """
        bboxes = get_box_tensor(bboxes)
        assert pred_bboxes.size(0) == bboxes.size(0)
        if pred_bboxes.ndim == 3:
            assert pred_bboxes.size(1) == bboxes.size(1)

        if pred_bboxes.ndim == 2 and not torch.onnx.is_in_onnx_export():
            # single image decode
            decoded_bboxes = delta2bbox(bboxes, pred_bboxes, self.means,
                                        self.stds, max_shape, wh_ratio_clip,
                                        self.clip_border, self.add_ctr_clamp,
                                        self.ctr_clamp)
        else:
            if pred_bboxes.ndim == 3 and not torch.onnx.is_in_onnx_export():
                warnings.warn(
                    'DeprecationWarning: onnx_delta2bbox is deprecated '
                    'in the case of batch decoding and non-ONNX, '
                    'please use “delta2bbox” instead. In order to improve '
                    'the decoding speed, the batch function will no '
                    'longer be supported. ')
            decoded_bboxes = onnx_delta2bbox(bboxes, pred_bboxes, self.means,
                                             self.stds, max_shape,
                                             wh_ratio_clip, self.clip_border,
                                             self.add_ctr_clamp,
                                             self.ctr_clamp)

        if self.use_box_type:
            assert decoded_bboxes.size(-1) == 4, \
                ('Cannot warp decoded boxes with box type when decoded boxes'
                 'have shape of (N, num_classes * 4)')
            decoded_bboxes = HorizontalBoxes(decoded_bboxes)
        return decoded_bboxes

View on GitHub (pinned to cfd5d3a985)

Solutions

Decode per-sample (loop over batch) or reshape so pred_bboxes is 2-D before decode
Ensure code uses delta2bbox path (the branch above) rather than falling into the ONNX helper
Update mmdet — newer implementations handle batched decoding natively

Example fix

# before
decoded = coder.decode(anchors, batched_pred_bboxes)  # ndim==3
# after
decoded = torch.cat([coder.decode(anchors[i:i+1].squeeze(0), batched_pred_bboxes[i]) for i in range(batched_pred_bboxes.shape[0])])

Defensive patterns

Strategy: validation

Validate before calling

assert pred_bboxes.ndim == 2 or torch.onnx.is_in_onnx_export(), 'decode batched deltas per-sample or use delta2bbox path'

Prevention

Keep decode inputs 2-D outside ONNX export
Batch: loop or vectorize with delta2bbox rather than the ONNX helper

When it happens

Trigger: Calling coder.decode(...) with pred_bboxes.ndim == 3 (batched deltas) while not inside torch.onnx.export; e.g. test-time batch prediction or aug-test code feeding batched proposals through a DeltaXYWHBBoxCoder on the non-ONNX path.

Common situations: Batch inference / multi-image test loaders hitting a decode path designed for ONNX tracing; upgrading mmdet versions where batch decoding via onnx_delta2bbox is being removed.

Related errors

AI-assisted analysis of open-mmlab/mmdetection@cfd5d3a985 (2026-08-27). Data as JSON: /api/errors/89d9cb4ed75df181. Report an issue: GitHub.