(beta) torch_npu.contrib.function.npu_bbox_coder_encode_xyxy2xywh

Supported Products

Product	Supported
Atlas A3 training products	√
Atlas A2 training products	√
Atlas inference products	√
Atlas training products	√

Function

Applies an NPU-based bounding box format encoding operation to convert format from xyxy to xywh.

Prototype

torch_npu.contrib.function.npu_bbox_coder_encode_xyxy2xywh(bboxes,gt_bboxes, means=None, stds=None, is_normalized=False, normalized_scale=10000.)

Parameters

bboxes (Tensor): Bounding boxes to be converted. This parameter must be 2D with shape (N, 4). The data type can be float or half.
gt_bboxes (Tensor): Ground truth bounding boxes used as a reference. This parameter must be 2D with shape (N, 4). The data type can be float or half.
means (List[float]): Optional. Mean used to denormalize the target delta coordinates. The default value is None.
stds (List[float]): Standard deviations used to denormalize delta coordinates. The default value is None.
is_normalized (bool): Indicates whether the coordinate values have been normalized. The default value is False.
normalized_scale (float): Normalization scale used to restore coordinates. The default value is 10000.

Return Values

Tensor

Bounding box transformation deltas.

Constraints

Dynamic shapes are not supported. Due to operator semantic limitations, only 2D scenarios with shape (N, 4) are supported. The shapes and data types of bboxes and gt_bboxes must be identical. The data type must be float16 or float32. The third input (stride) must be a 1D tensor, and its first dimension must match that of the first input (bboxes).

Example

>>> import torch, torch_npu
>>> from torch_npu.contrib.function import npu_bbox_coder_encode_xyxy2xywh
>>> A = 1024
>>> bboxes = torch.randint(0, 512, size=(A, 4)).float().npu()
>>> gt_bboxes = torch.randint(0, 512, size=(A, 4)).float().npu()
>>> out = npu_bbox_coder_encode_xyxy2xywh(bboxes, gt_bboxes)
>>> torch.npu.synchronize()
>>> print('npu_bbox_coder_encode_xyxy2xywh done. output shape is ', out.shape)
npu_bbox_coder_encode_xyxy2xywh done. output shape is torch.Size([1024, 4])