feat: 迁移Yolo系列和TransArgb ONNX插件-ops-cv-AtomGit

已合并

feat: 迁移Yolo系列和TransArgb ONNX插件 #1122

tianqiguang创建于 7月14日从tianqiguang/ops-cv： master 合入到 cann/ops-cv： master

feat: 迁移Yolo系列和TransArgb ONNX插件 #1122

已合并

tianqiguang创建于 7月14日

从tianqiguang/ops-cv master合入到cann/ops-cv master

共 9 个文件变更+968-229

M.clang-format+4-0

	@@ -47,6 +47,10 @@ AllowAllParametersOfDeclarationOnNextLine: true
47	AlwaysBreakBeforeMultilineStrings: false	47	AlwaysBreakBeforeMultilineStrings: false
48	AlwaysBreakTemplateDeclarations: true	48	AlwaysBreakTemplateDeclarations: true
49		49
		50	+# 宏语句定义
		51	+StatementMacros:
		52	+ - OP_END_FACTORY_REG
		53	+
50	BinPackArguments: true	54	BinPackArguments: true
51	BreakBeforeTernaryOperators: false	55	BreakBeforeTernaryOperators: false
52	SpacesBeforeTrailingComments: 1	56	SpacesBeforeTrailingComments: 1

Mcommon/inc/op_graph/op_cv_proto_extend.h+469-229

	@@ -1,230 +1,470 @@
1	-/**	1	+/**
2	- * Copyright (c) 2025 Huawei Technologies Co., Ltd.	2	+ * Copyright (c) 2025 Huawei Technologies Co., Ltd.
3	- * This program is free software, you can redistribute it and/or modify it under the terms and conditions of	3	+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4	- * CANN Open Software License Agreement Version 2.0 (the "License").	4	+ * CANN Open Software License Agreement Version 2.0 (the "License").
5	- * Please refer to the License for details. You may not use this file except in compliance with the License.	5	+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6	- * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,	6	+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7	- * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.	7	+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8	- * See LICENSE in the root of the software repository for the full text of the License.	8	+ * See LICENSE in the root of the software repository for the full text of the License.
9	- */	9	+ */
10	-	10	+
11	-/*!	11	+/*!
12	- * \file op_cv_proto_extend.h	12	+ * \file op_cv_proto_extend.h
13	- * \brief	13	+ * \brief
14	- */	14	+ */
15	-#ifndef OPS_OP_CV_PROTO_EXTEND_H_	15	+#ifndef OPS_OP_CV_PROTO_EXTEND_H_
16	-#define OPS_OP_CV_PROTO_EXTEND_H_	16	+#define OPS_OP_CV_PROTO_EXTEND_H_
17	-	17	+
18	-#include "graph/operator_reg.h"	18	+#include "graph/operator_reg.h"
19	-	19	+
20	-namespace ge {	20	+namespace ge {
21	-/**	21	+/**
22	-* @brief Performs Col2ImV2 for each batch entry. \n	22	+* @brief Performs Col2ImV2 for each batch entry. \n
23	-	23	+
24	-* @par Inputs:	24	+* @par Inputs:
25	-* @li x: The Col Tensor. 3-D, shape: `(n, ckernel_hkernel_w, ho*wo)`.	25	+* @li x: The Col Tensor. 3-D, shape: `(n, ckernel_hkernel_w, ho*wo)`.
26	-where ho/wo is do = (output_d + 2padding_d - dilation_d(kernel_d - 1) - 1)//stride_d + 1.	26	+where ho/wo is do = (output_d + 2padding_d - dilation_d(kernel_d - 1) - 1)//stride_d + 1.
27	-* @li output_size: The img shape Tensor. 1-D, shape:`(2)`, value: (output_h, output_w).	27	+* @li output_size: The img shape Tensor. 1-D, shape:`(2)`, value: (output_h, output_w).
28	-* @li kernel_shape: The kernel size Tensor. 1-D , value: `(kernel_h, kernel_w)`, the shape of kernel in convolution. \n	28	+* @li kernel_shape: The kernel size Tensor. 1-D , value: `(kernel_h, kernel_w)`, the shape of kernel in convolution. \n
29	-	29	+
30	-* @par Outputs:	30	+* @par Outputs:
31	-* y: The img Tensor. 4-D, shape: `(n, c, output_h, output_w)`. \n	31	+* y: The img Tensor. 4-D, shape: `(n, c, output_h, output_w)`. \n
32	-	32	+
33	-* @par Attributes:	33	+* @par Attributes:
34	-	34	+
35	-* @li dilation: ListInt, value: `(dilation_h, dilation_w)`, the dilation in convolution.	35	+* @li dilation: ListInt, value: `(dilation_h, dilation_w)`, the dilation in convolution.
36	-* @li padding: ListInt, value: `(padding_h, padding_w)`, the dilation in convolution.	36	+* @li padding: ListInt, value: `(padding_h, padding_w)`, the dilation in convolution.
37	-* @li stride: ListInt, value: `(stride_h, stride_w)`, the dilation in convolution. \n	37	+* @li stride: ListInt, value: `(stride_h, stride_w)`, the dilation in convolution. \n
38	-	38	+
39	-* @par Third-party framework compatibility	39	+* @par Third-party framework compatibility
40	-* Compatible with ONNX Col2Im operator.	40	+* Compatible with ONNX Col2Im operator.
41	-*/	41	+*/
42	-REG_OP(Col2ImV2)	42	+REG_OP(Col2ImV2)
43	- .INPUT(x, TensorType({DT_FLOAT, DT_FLOAT16}))	43	+ .INPUT(x, TensorType({DT_FLOAT, DT_FLOAT16}))
44	- .INPUT(output_size, TensorType({DT_INT32, DT_INT32}))	44	+ .INPUT(output_size, TensorType({DT_INT32, DT_INT32}))
45	- .INPUT(kernel_size, TensorType({DT_INT32, DT_INT32}))	45	+ .INPUT(kernel_size, TensorType({DT_INT32, DT_INT32}))
46	- .OUTPUT(y, TensorType({DT_FLOAT, DT_FLOAT16}))	46	+ .OUTPUT(y, TensorType({DT_FLOAT, DT_FLOAT16}))
47	- .REQUIRED_ATTR(dilation, ListInt)	47	+ .REQUIRED_ATTR(dilation, ListInt)
48	- .REQUIRED_ATTR(padding, ListInt)	48	+ .REQUIRED_ATTR(padding, ListInt)
49	- .REQUIRED_ATTR(stride, ListInt)	49	+ .REQUIRED_ATTR(stride, ListInt)
50	- .OP_END_FACTORY_REG(Col2ImV2)	50	+ .OP_END_FACTORY_REG(Col2ImV2)
51	-	51	+
52	- /**	52	+/**
53	- * @brief First calculate the minimum closure area of the two boxes, IoU,	53	+* @brief First calculate the minimum closure area of the two boxes, IoU,
54	- * the proportion of the closed area that does not belong to the two boxes in the closure area,	54	+* the proportion of the closed area that does not belong to the two boxes in the closure area,
55	- * and finally subtract this proportion from IoU to get GIoU . \n	55	+* and finally subtract this proportion from IoU to get GIoU . \n
56	-	56	+
57	- * @par Inputs:	57	+* @par Inputs:
58	- * Two inputs, including:	58	+* Two inputs, including:
59	- * @li bboxes: Bounding boxes, a 2D Tensor of type float16 or float32 with	59	+* @li bboxes: Bounding boxes, a 2D Tensor of type float16 or float32 with
60	- * shape (N, 4). "N" indicates the number of bounding boxes, and the value	60	+* shape (N, 4). "N" indicates the number of bounding boxes, and the value
61	- * "4" refers to [x1, y1, x2, y2] or [x, y, w, h].	61	+* "4" refers to [x1, y1, x2, y2] or [x, y, w, h].
62	- * @li gtboxes: Ground-truth boxes, a 2D Tensor of type float16 or float32	62	+* @li gtboxes: Ground-truth boxes, a 2D Tensor of type float16 or float32
63	- * with shape (M, 4). "M" indicates the number of ground truth boxes, and	63	+* with shape (M, 4). "M" indicates the number of ground truth boxes, and
64	- * the value "4" refers to [x1, y1, x2, y2] or [x, y, w, h] . \n	64	+* the value "4" refers to [x1, y1, x2, y2] or [x, y, w, h] . \n
65	-	65	+
66	- * @par Attributes:	66	+* @par Attributes:
67	- * @li trans: An optional bool, true for 'xywh', false for 'xyxy'.	67	+* @li trans: An optional bool, true for 'xywh', false for 'xyxy'.
68	- * @li is_cross: An optional bool, control whether the output shape is [M, N] or [1, N]	68	+* @li is_cross: An optional bool, control whether the output shape is [M, N] or [1, N]
69	- * @li mode: Computation mode, a character string with the value range of [iou, iof] . \n	69	+* @li mode: Computation mode, a character string with the value range of [iou, iof] . \n
70	-	70	+
71	- * @par Outputs:	71	+* @par Outputs:
72	- * overlap: A 2D Tensor of type float16 or float32 with shape [M, N] or [1, N],	72	+* overlap: A 2D Tensor of type float16 or float32 with shape [M, N] or [1, N],
73	- * specifying the IoU or IoF ratio . \n	73	+* specifying the IoU or IoF ratio . \n
74	-	74	+
75	- * @attention Constraints:	75	+* @attention Constraints:
76	- * Only computation of float16 data is supported. To avoid overflow, the input	76	+* Only computation of float16 data is supported. To avoid overflow, the input
77	- * length and width are scaled by 0.2 internally.	77	+* length and width are scaled by 0.2 internally.
78	- */	78	+*/
79	- REG_OP(GIoU)	79	+REG_OP(GIoU)
80	- .INPUT(bboxes, TensorType({DT_FLOAT16, DT_FLOAT}))	80	+ .INPUT(bboxes, TensorType({DT_FLOAT16, DT_FLOAT}))
81	- .INPUT(gtboxes, TensorType({DT_FLOAT16, DT_FLOAT}))	81	+ .INPUT(gtboxes, TensorType({DT_FLOAT16, DT_FLOAT}))
82	- .OUTPUT(overlap, TensorType({DT_FLOAT16, DT_FLOAT}))	82	+ .OUTPUT(overlap, TensorType({DT_FLOAT16, DT_FLOAT}))
83	- .ATTR(trans, Bool, false)	83	+ .ATTR(trans, Bool, false)
84	- .ATTR(is_cross, Bool, true)	84	+ .ATTR(is_cross, Bool, true)
85	- .ATTR(mode, String, "iou")	85	+ .ATTR(mode, String, "iou")
86	- .OP_END_FACTORY_REG(GIoU)	86	+ .OP_END_FACTORY_REG(GIoU)
87	-	87	+
88	- /**	88	+/**
89	- *@brief RotatedIou . \n	89	+*@brief RotatedIou . \n
90	-	90	+
91	- * @par Inputs:	91	+* @par Inputs:
92	- *@li boxes : data of grad increment, a 3D Tensor of type float32 with	92	+*@li boxes : data of grad increment, a 3D Tensor of type float32 with
93	- * shape (B, 5, N). "N" indicates the number of boxes, and the value	93	+* shape (B, 5, N). "N" indicates the number of boxes, and the value
94	- * "5" refers to [x1, y1, x2, y2, theta] or [x, y, w, h, theta].	94	+* "5" refers to [x1, y1, x2, y2, theta] or [x, y, w, h, theta].
95	- * @li query_boxes: Bounding boxes, a 3D Tensor of type float32 with	95	+* @li query_boxes: Bounding boxes, a 3D Tensor of type float32 with
96	- * shape (B, 5, K). "K" indicates the number of boxes, and the value	96	+* shape (B, 5, K). "K" indicates the number of boxes, and the value
97	- * "5" refers to [x1, y1, x2, y2, theta] or [x, y, w, h, theta].	97	+* "5" refers to [x1, y1, x2, y2, theta] or [x, y, w, h, theta].
98	-	98	+
99	- * @par Attributes:	99	+* @par Attributes:
100	- *@li trans: An optional attr, true for 'xyxyt', false for 'xywht'.	100	+*@li trans: An optional attr, true for 'xyxyt', false for 'xywht'.
101	- * @li mode: An optional attr, a character string with the value range of ['iou', 'iof'],	101	+* @li mode: An optional attr, a character string with the value range of ['iou', 'iof'],
102	- * only support 'iou' now.	102	+* only support 'iou' now.
103	- *@li is_cross: Cross calculation when it is True, and one-to-one calculation when it is False.	103	+*@li is_cross: Cross calculation when it is True, and one-to-one calculation when it is False.
104	- *@li v_threshold: An optional attr, provide condition relaxation for intersection calculation.	104	+*@li v_threshold: An optional attr, provide condition relaxation for intersection calculation.
105	- * @li e_threshold: An optional attr, provide condition relaxation for intersection calculation.	105	+* @li e_threshold: An optional attr, provide condition relaxation for intersection calculation.
106	-	106	+
107	- *@par Outputs:	107	+*@par Outputs:
108	- * iou: A 3D Tensor of float32 with shape [B, N, K].	108	+* iou: A 3D Tensor of float32 with shape [B, N, K].
109	-	109	+
110	- *@attention Constraints:	110	+*@attention Constraints:
111	- * In each batch, the invalid box cannot appear before the valid box.	111	+* In each batch, the invalid box cannot appear before the valid box.
112	- */	112	+*/
113	- REG_OP(RotatedIou)	113	+REG_OP(RotatedIou)
114	- .INPUT(boxes, TensorType({DT_FLOAT}))	114	+ .INPUT(boxes, TensorType({DT_FLOAT}))
115	- .INPUT(query_boxes, TensorType({DT_FLOAT}))	115	+ .INPUT(query_boxes, TensorType({DT_FLOAT}))
116	- .OUTPUT(iou, TensorType({DT_FLOAT}))	116	+ .OUTPUT(iou, TensorType({DT_FLOAT}))
117	- .ATTR(trans, Bool, false)	117	+ .ATTR(trans, Bool, false)
118	- .ATTR(mode, String, "iou")	118	+ .ATTR(mode, String, "iou")
119	- .ATTR(is_cross, Bool, true)	119	+ .ATTR(is_cross, Bool, true)
120	- .ATTR(v_threshold, Float, 0)	120	+ .ATTR(v_threshold, Float, 0)
121	- .ATTR(e_threshold, Float, 0)	121	+ .ATTR(e_threshold, Float, 0)
122	- .OP_END_FACTORY_REG(RotatedIou)	122	+ .OP_END_FACTORY_REG(RotatedIou)
123	-	123	+
124	- /**	124	+/**
125	- * @brief Resize the input tensor. \n	125	+* @brief Resize the input tensor. \n
126	- currently, only support resize image tensor using nearest neighbor and linear interpolation.	126	+currently, only support resize image tensor using nearest neighbor and linear interpolation.
127	-	127	+
128	- * @par Inputs:	128	+* @par Inputs:
129	- * Input x must be a 4-D tensor. Inputs include: \n	129	+* Input x must be a 4-D tensor. Inputs include: \n
130	- * @li x: A Tensor. Must be one of the following types: uint8, int8, int16, \n	130	+* @li x: A Tensor. Must be one of the following types: uint8, int8, int16, \n
131	- int32, int64, float16, float, double. 4-D with shape [batch, height, width, channels] \n	131	+int32, int64, float16, float, double. 4-D with shape [batch, height, width, channels] \n
132	- or shape [batch, channels, height, width].	132	+or shape [batch, channels, height, width].
133	- * @li roi: A 1-D float Tensor. Only takes effect when attr coordinate_transformation_mode \n	133	+* @li roi: A 1-D float Tensor. Only takes effect when attr coordinate_transformation_mode \n
134	- is "tf_crop_and_resize". Must be one of the following types: float16, float, double.	134	+is "tf_crop_and_resize". Must be one of the following types: float16, float, double.
135	- * @li scales: A 1-D float Tensor, the scale array along each dimension, Only one of \n	135	+* @li scales: A 1-D float Tensor, the scale array along each dimension, Only one of \n
136	- 'scales' and 'sizes' can be specified. Must be float type.	136	+'scales' and 'sizes' can be specified. Must be float type.
137	- * @li sizes: A 1-D int64 Tensor, The size of the output tensor. Only one of \n	137	+* @li sizes: A 1-D int64 Tensor, The size of the output tensor. Only one of \n
138	- 'scales' and 'sizes' can be specified. If 'size' is specified, then set scales \n	138	+'scales' and 'sizes' can be specified. If 'size' is specified, then set scales \n
139	- to empty data (zero shape) in this operator's input list. Must be one of \n	139	+to empty data (zero shape) in this operator's input list. Must be one of \n
140	- the following types: int32, int64.	140	+the following types: int32, int64.
141	-	141	+
142	- * @par Attributes:	142	+* @par Attributes:
143	- * @li coordinate_transformation_mode: An optional String. how to transform \n	143	+* @li coordinate_transformation_mode: An optional String. how to transform \n
144	- the coordinate in the resized tensor to the coordinate in the original tensor. \n	144	+the coordinate in the resized tensor to the coordinate in the original tensor. \n
145	- options: pytorch_half_pixel, align_corners, asymmetric, \n	145	+options: pytorch_half_pixel, align_corners, asymmetric, \n
146	- tf_crop_and_resize.	146	+tf_crop_and_resize.
147	- * @li cubic_coeff_a: An optional Float. Defaults to -0.75, only used in cubic interpolation. \n	147	+* @li cubic_coeff_a: An optional Float. Defaults to -0.75, only used in cubic interpolation. \n
148	- other optional: -0.5	148	+other optional: -0.5
149	- * @li exclude_outside: An optional Int. Defaults to 0, If set to 1, the weight of sampling \n	149	+* @li exclude_outside: An optional Int. Defaults to 0, If set to 1, the weight of sampling \n
150	- locations outside the tensor will be set to 0 and the weight will be renormalized \n	150	+locations outside the tensor will be set to 0 and the weight will be renormalized \n
151	- so that their sum is 1.0.	151	+so that their sum is 1.0.
152	- * @li extrapolation_value: An optional Float. Defaults to 0.0f. When coordinate_transformation_mode \n	152	+* @li extrapolation_value: An optional Float. Defaults to 0.0f. When coordinate_transformation_mode \n
153	- is "tf_crop_and_resize" and x_original is outside the range [0, length_original - 1], \n	153	+is "tf_crop_and_resize" and x_original is outside the range [0, length_original - 1], \n
154	- this value is used as the corresponding output value.	154	+this value is used as the corresponding output value.
155	- * @li mode: An optional String. Defaults to nearest. Three interpolation modes: nearest (default), \n	155	+* @li mode: An optional String. Defaults to nearest. Three interpolation modes: nearest (default), \n
156	- linear and cubic.	156	+linear and cubic.
157	- * @li nearest_mode: An optional String. Defaults to round_prefer_floor. Four modes: round_prefer_floor, \n	157	+* @li nearest_mode: An optional String. Defaults to round_prefer_floor. Four modes: round_prefer_floor, \n
158	- round_prefer_ceil, floor, ceil. Only used by nearest interpolation.	158	+round_prefer_ceil, floor, ceil. Only used by nearest interpolation.
159	-	159	+
160	- * @par Outputs:	160	+* @par Outputs:
161	- * y: A Tensor. Has the same type as x.	161	+* y: A Tensor. Has the same type as x.
162	-	162	+
163	- * @attention Constraints: \n	163	+* @attention Constraints: \n
164	- * Input x must be a 4-D tensor.	164	+* Input x must be a 4-D tensor.
165	-	165	+
166	- * @par Third-party framework compatibility	166	+* @par Third-party framework compatibility
167	- * Compatible with tensorflow ResizeNearestNeighborV2 operator.	167	+* Compatible with tensorflow ResizeNearestNeighborV2 operator.
168	- */	168	+*/
169	-	169	+
170	- REG_OP(Resize)	170	+REG_OP(Resize)
171	- .INPUT(x, TensorType({DT_INT8, DT_UINT8, DT_INT16, DT_UINT16, DT_INT32, DT_INT64, DT_FLOAT16, DT_FLOAT, DT_DOUBLE}))	171	+ .INPUT(x, TensorType({DT_INT8, DT_UINT8, DT_INT16, DT_UINT16, DT_INT32, DT_INT64, DT_FLOAT16, DT_FLOAT, DT_DOUBLE}))
172	- .OPTIONAL_INPUT(roi, TensorType({DT_FLOAT16, DT_FLOAT, DT_DOUBLE}))	172	+ .OPTIONAL_INPUT(roi, TensorType({DT_FLOAT16, DT_FLOAT, DT_DOUBLE}))
173	- .OPTIONAL_INPUT(scales, TensorType({DT_FLOAT}))	173	+ .OPTIONAL_INPUT(scales, TensorType({DT_FLOAT}))
174	- .OPTIONAL_INPUT(sizes, TensorType({DT_INT64, DT_INT32}))	174	+ .OPTIONAL_INPUT(sizes, TensorType({DT_INT64, DT_INT32}))
175	- .OUTPUT(y,	175	+ .OUTPUT(y,
176	- TensorType({DT_INT8, DT_UINT8, DT_INT16, DT_UINT16, DT_INT32, DT_INT64, DT_FLOAT16, DT_FLOAT, DT_DOUBLE}))	176	+ TensorType({DT_INT8, DT_UINT8, DT_INT16, DT_UINT16, DT_INT32, DT_INT64, DT_FLOAT16, DT_FLOAT, DT_DOUBLE}))
177	- .ATTR(coordinate_transformation_mode, String, "half_pixel")	177	+ .ATTR(coordinate_transformation_mode, String, "half_pixel")
178	- .ATTR(cubic_coeff_a, Float, -0.75)	178	+ .ATTR(cubic_coeff_a, Float, -0.75)
179	- .ATTR(exclude_outside, Int, 0)	179	+ .ATTR(exclude_outside, Int, 0)
180	- .ATTR(extrapolation_value, Float, 0.0)	180	+ .ATTR(extrapolation_value, Float, 0.0)
181	- .ATTR(mode, String, "nearest")	181	+ .ATTR(mode, String, "nearest")
182	- .ATTR(nearest_mode, String, "round_prefer_floor")	182	+ .ATTR(nearest_mode, String, "round_prefer_floor")
183	- .OP_END_FACTORY_REG(Resize)	183	+ .OP_END_FACTORY_REG(Resize)
184	-	184	+
185	- /**	185	+/**
186	- *@brief Greedily selects a subset of bounding boxes in descending order of	186	+*@brief Greedily selects a subset of bounding boxes in descending order of
187	- score . \n	187	+score . \n
188	-	188	+
189	- *@par Inputs:	189	+*@par Inputs:
190	- *Input boxes and scores must be float type. Inputs include:	190	+*Input boxes and scores must be float type. Inputs include:
191	- *@li boxes: A 2-D float tensor of shape [num_boxes, 4].	191	+*@li boxes: A 2-D float tensor of shape [num_boxes, 4].
192	- *@li scores: A 1-D float tensor of shape [num_boxes] representing a single	192	+*@li scores: A 1-D float tensor of shape [num_boxes] representing a single
193	- score corresponding to each box (each row of boxes).	193	+score corresponding to each box (each row of boxes).
194	- *@li max_output_size: A scalar integer tensor representing the maximum number	194	+*@li max_output_size: A scalar integer tensor representing the maximum number
195	- of boxes to be selected by non max suppression.	195	+of boxes to be selected by non max suppression.
196	- *@li iou_threshold: A 0-D float tensor representing the threshold for deciding	196	+*@li iou_threshold: A 0-D float tensor representing the threshold for deciding
197	- whether boxes overlap too much with respect to IOU.	197	+whether boxes overlap too much with respect to IOU.
198	- *@li score_threshold: A 0-D float tensor representing the threshold for	198	+*@li score_threshold: A 0-D float tensor representing the threshold for
199	- deciding when to remove boxes based on score . \n	199	+deciding when to remove boxes based on score . \n
200	-	200	+
201	- *@par Attributes:	201	+*@par Attributes:
202	- *pad_to_max_output_size: If true, the output selected_indices is padded	202	+*pad_to_max_output_size: If true, the output selected_indices is padded
203	- to be of length max_output_size. Defaults to false . \n	203	+to be of length max_output_size. Defaults to false . \n
204	-	204	+
205	- *@par Outputs:	205	+*@par Outputs:
206	- *@li selected_indices: A 1-D integer tensor of shape [M] representing the	206	+*@li selected_indices: A 1-D integer tensor of shape [M] representing the
207	- selected indices from the boxes tensor, where M <= max_output_size.	207	+selected indices from the boxes tensor, where M <= max_output_size.
208	- *@li valid_outputs: A 0-D integer tensor representing the number of valid	208	+*@li valid_outputs: A 0-D integer tensor representing the number of valid
209	- elements in selected_indices, with the valid elements appearing first . \n	209	+elements in selected_indices, with the valid elements appearing first . \n
210	-	210	+
211	- *@attention Constraints:	211	+*@attention Constraints:
212	- *Input boxes and scores must be float type . \n	212	+*Input boxes and scores must be float type . \n
213	-	213	+
214	- *@par Third-party framework compatibility	214	+*@par Third-party framework compatibility
215	- *Compatible with tensorflow NonMaxSuppressionV4 operator.	215	+*Compatible with tensorflow NonMaxSuppressionV4 operator.
216	- */	216	+*/
217	-	217	+
218	- REG_OP(NonMaxSuppressionV4)	218	+REG_OP(NonMaxSuppressionV4)
219	- .INPUT(boxes, TensorType({DT_FLOAT16, DT_FLOAT}))	219	+ .INPUT(boxes, TensorType({DT_FLOAT16, DT_FLOAT}))
220	- .INPUT(scores, TensorType({DT_FLOAT16, DT_FLOAT}))	220	+ .INPUT(scores, TensorType({DT_FLOAT16, DT_FLOAT}))
221	- .INPUT(max_output_size, TensorType({DT_INT32}))	221	+ .INPUT(max_output_size, TensorType({DT_INT32}))
222	- .INPUT(iou_threshold, TensorType({DT_FLOAT16, DT_FLOAT}))	222	+ .INPUT(iou_threshold, TensorType({DT_FLOAT16, DT_FLOAT}))
223	- .INPUT(score_threshold, TensorType({DT_FLOAT16, DT_FLOAT}))	223	+ .INPUT(score_threshold, TensorType({DT_FLOAT16, DT_FLOAT}))
224	- .OUTPUT(selected_indices, TensorType({DT_INT32}))	224	+ .OUTPUT(selected_indices, TensorType({DT_INT32}))
225	- .OUTPUT(valid_outputs, TensorType({DT_INT32}))	225	+ .OUTPUT(valid_outputs, TensorType({DT_INT32}))
226	- .ATTR(pad_to_max_output_size, Bool, false)	226	+ .ATTR(pad_to_max_output_size, Bool, false)
227	- .OP_END_FACTORY_REG(NonMaxSuppressionV4)	227	+ .OP_END_FACTORY_REG(NonMaxSuppressionV4)
228	-} // namespace ge	228	+
229	-	229	+/**
		230	+*@brief Normalizes data. It is called Region on YOLO v2 and Yolo on YOLO v3 . \n
		231	+
		232	+*@par Inputs:
		233	+*x: An NCHW tensor of type float16 or float32. The data is with shape (N,
		234	+* boxes*(coords+obj+classes), H, W) . \n
		235	+
		236	+*@par Attributes:
		237	+*@li boxes: An optional int32, specifying the number of anchor boxes. Defaults to "3".
		238	+*@li coords: An int32, specifying the number of parameters required for locating an object. Defaults to "4".
		239	+*@li classes: An int32, specifying the number of prediction classes. Defaults to "80".
		240	+*@li yolo_version: A string, specifying the YOLO version, either "V2" or "V3". Defaults to "V3"
		241	+*@li softmax: A bool, specifying whether to perform softmax. Defaults to "false".
		242	+*@li background: A bool. Defaults to "false".
		243	+*@li softmaxtree: A bool. Defaults to "false" . \n
		244	+
		245	+*@par Outputs:
		246	+*@li coord_data: Specifies the coordinates of a detected box.
		247	+*@li obj_prob: Specifies the confidence.
		248	+*@li classes_prob: Specifies the prediction classes . \n
		249	+*/
		250	+#ifndef OPS_PROTO_DEF_YOLO
		251	+#define OPS_PROTO_DEF_YOLO
		252	+REG_OP(Yolo)
		253	+ .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT}))
		254	+ .OUTPUT(coord_data, TensorType({DT_FLOAT16, DT_FLOAT}))
		255	+ .OUTPUT(obj_prob, TensorType({DT_FLOAT16, DT_FLOAT}))
		256	+ .OUTPUT(classes_prob, TensorType({DT_FLOAT16, DT_FLOAT}))
		257	+ .ATTR(boxes, Int, 3)
		258	+ .ATTR(coords, Int, 4)
		259	+ .ATTR(classes, Int, 80)
		260	+ .ATTR(yolo_version, String, "V3")
		261	+ .ATTR(softmax, Bool, false)
		262	+ .ATTR(background, Bool, false)
		263	+ .ATTR(softmaxtree, Bool, false)
		264	+ .OP_END_FACTORY_REG(Yolo)
		265	+#endif
		266	+
		267	+/**
		268	+*@brief Normalizes data. It is called Region on YOLO v2 and Yolo on YOLO v3 . \n
		269	+
		270	+*@par Inputs:
		271	+*x: An NCHW tensor of type float16 or float32 . \n
		272	+
		273	+*@par Attributes:
		274	+*@li boxes: An optional int32, specifying the number of anchor boxes. Defaults to "3".
		275	+*@li coords: An int32, specifying the number of parameters required for locating an object. Defaults to "4".
		276	+*@li classes: An int32, specifying the number of prediction classes. Defaults to "80".
		277	+*@li yolo_version: A string, specifying the YOLO version. Defaults to "V5"
		278	+*@li softmax: A bool, specifying whether to perform softmax. Defaults to "false".
		279	+*@li background: A bool. Defaults to "false".
		280	+*@li softmaxtree: A bool. Defaults to "false" . \n
		281	+
		282	+*@par Outputs:
		283	+*@li coord_data: Specifies the coordinates of a detected box.
		284	+*@li obj_prob: Specifies the confidence.
		285	+*@li classes_prob: Specifies the prediction classes . \n
		286	+*/
		287	+#ifndef OPS_PROTO_DEF_YOLOPREDETECTION
		288	+#define OPS_PROTO_DEF_YOLOPREDETECTION
		289	+REG_OP(YoloPreDetection)
		290	+ .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT}))
		291	+ .OUTPUT(coord_data, TensorType({DT_FLOAT16, DT_FLOAT}))
		292	+ .OUTPUT(obj_prob, TensorType({DT_FLOAT16, DT_FLOAT}))
		293	+ .OUTPUT(classes_prob, TensorType({DT_FLOAT16, DT_FLOAT}))
		294	+ .ATTR(boxes, Int, 3)
		295	+ .ATTR(coords, Int, 4)
		296	+ .ATTR(classes, Int, 80)
		297	+ .ATTR(yolo_version, String, "V5")
		298	+ .ATTR(softmax, Bool, false)
		299	+ .ATTR(background, Bool, false)
		300	+ .ATTR(softmaxtree, Bool, false)
		301	+ .OP_END_FACTORY_REG(YoloPreDetection)
		302	+#endif
		303	+
		304	+/**
		305	+*@brief Performs YOLO V5 detection . \n
		306	+
		307	+*@par Inputs:
		308	+*@li x: Dynamic inputs, the outputs of operator Yolo at the preceding layer.
		309	+*@li img_info: A float16 or float32, describing the image information . \n
		310	+
		311	+*@par Attributes:
		312	+*@li biases: A required float. "biases = Number of Yolo operators at the preceding layer x 2 x boxes"
		313	+*@li boxes: A required int32, specifying the number of anchor boxes predicted for each Yolo layer.
		314	+*@li coords: Specifies the number of coordinate parameters. Must be 4.
		315	+*@li classes: A required int32, specifying the number of classes to be predicted.
		316	+*@li relative: An optional bool. Defaults to and must be "true".
		317	+*@li obj_threshold: A required float, specifying the confidence threshold for box filtering.
		318	+*@li post_nms_topn: An optional int32. This attribute is reserved.
		319	+*@li score_threshold: A required float, specifying the class score threshold for box filtering.
		320	+*@li iou_threshold: A required float, specifying the intersection-over-union (IOU) threshold for box filtering.\n
		321	+*@li pre_nms_topn: An optional int, specifying the number of boxes for non-maximum suppression (NMS). Defaults to "512".
		322	+*@li N: An optional int. Defaults to "10".
		323	+*@li resize_origin_img_to_net: An optional bool. Defaults to "false".
		324	+*@li out_box_dim: An optional int. Defaults to "3".
		325	+*@li alpha: An optional float. Defaults to "2.0".
		326	+
		327	+*@par Outputs:
		328	+*@li box_out: A tensor of type float16 or float32, describing the information of each output box.
		329	+*@li box_out_num: A tensor of type int32, specifying the number of output boxes . \n
		330	+*/
		331	+#ifndef OPS_PROTO_DEF_YOLOV5DETECTIONOUTPUT
		332	+#define OPS_PROTO_DEF_YOLOV5DETECTIONOUTPUT
		333	+REG_OP(YoloV5DetectionOutput)
		334	+ .DYNAMIC_INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT}))
		335	+ .REQUIRED_ATTR(biases, ListFloat)
		336	+ .ATTR(boxes, Int, 3)
		337	+ .ATTR(coords, Int, 4)
		338	+ .ATTR(classes, Int, 80)
		339	+ .ATTR(relative, Bool, true)
		340	+ .ATTR(obj_threshold, Float, 0.5)
		341	+ .ATTR(post_nms_topn, Int, 512)
		342	+ .ATTR(score_threshold, Float, 0.5)
		343	+ .ATTR(iou_threshold, Float, 0.45f)
		344	+ .ATTR(pre_nms_topn, Int, 512)
		345	+ .ATTR(N, Int, 10)
		346	+ .ATTR(resize_origin_img_to_net, Bool, false)
		347	+ .ATTR(out_box_dim, Int, 3)
		348	+ .ATTR(alpha, Float, 2.0)
		349	+ .OUTPUT(box_out, TensorType({DT_FLOAT16, DT_FLOAT}))
		350	+ .OUTPUT(box_out_num, TensorType({DT_INT32}))
		351	+ .OP_END_FACTORY_REG(YoloV5DetectionOutput)
		352	+#endif
		353	+
		354	+/**
		355	+*@brief Performs YOLO V3 detection output V2 . \n
		356	+
		357	+*@par Inputs:
		358	+*@li x: Dynamic inputs, the outputs of operator Yolo at the preceding layer . \n
		359	+
		360	+*@par Attributes:
		361	+*@li biases: A float list. "biases = Number of Yolo operators at the preceding layer x 2 x boxes"
		362	+*@li boxes: A required int, specifying the number of anchor boxes predicted for each Yolo layer.
		363	+*@li coords: Specifies the number of coordinate parameters. Must be 4.
		364	+*@li classes: A required int, specifying the number of classes to be predicted.
		365	+*@li relative: An optional bool. Defaults to and must be "true".
		366	+*@li obj_threshold: A required float, specifying the confidence threshold for box filtering.
		367	+*@li post_nms_topn: An optional int. This attribute is reserved.
		368	+*@li score_threshold: A required float, specifying the class score threshold for box filtering.
		369	+*@li iou_threshold: A required float, specifying the intersection-over-union (IOU) threshold for box filtering.
		370	+*@li pre_nms_topn: An optional int, specifying the number of boxes for non-maximum suppression (NMS). Defaults to "512".
		371	+*@li N: An optional int. Defaults to "10".
		372	+*@li resize_origin_img_to_net: An optional bool. Defaults to "false".
		373	+*@li out_box_dim: An optional int. Defaults to "3".
		374	+
		375	+*@par Outputs:
		376	+*@li box_out: A tensor of type float16 or float32, describing the information of each output box.
		377	+*@li box_out_num: A tensor of type int32, specifying the number of output boxes . \n
		378	+*/
		379	+#ifndef OPS_PROTO_DEF_YOLOV3DETECTIONOUTPUTV2
		380	+#define OPS_PROTO_DEF_YOLOV3DETECTIONOUTPUTV2
		381	+REG_OP(YoloV3DetectionOutputV2)
		382	+ .DYNAMIC_INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT}))
		383	+ .REQUIRED_ATTR(biases, ListFloat)
		384	+ .ATTR(boxes, Int, 3)
		385	+ .ATTR(coords, Int, 4)
		386	+ .ATTR(classes, Int, 80)
		387	+ .ATTR(relative, Bool, true)
		388	+ .ATTR(obj_threshold, Float, 0.5)
		389	+ .ATTR(post_nms_topn, Int, 512)
		390	+ .ATTR(score_threshold, Float, 0.5)
		391	+ .ATTR(iou_threshold, Float, 0.45f)
		392	+ .ATTR(pre_nms_topn, Int, 512)
		393	+ .ATTR(N, Int, 10)
		394	+ .ATTR(resize_origin_img_to_net, Bool, false)
		395	+ .ATTR(out_box_dim, Int, 3)
		396	+ .OUTPUT(box_out, TensorType({DT_FLOAT16, DT_FLOAT}))
		397	+ .OUTPUT(box_out_num, TensorType({DT_INT32}))
		398	+ .OP_END_FACTORY_REG(YoloV3DetectionOutputV2)
		399	+#endif
		400	+
		401	+/**
		402	+*@brief Generates bounding boxes based on "anchor_boxes" and "gt_bboxes" . \n
		403	+
		404	+*@par Inputs:
		405	+*@li anchor_boxes: anchor sample boxes of origin image
		406	+*@li gt_bboxes: ground truth boxes
		407	+*@li stride: stride . \n
		408	+
		409	+*@par Attributes:
		410	+*performance_mode: select performance mode, "high_precision" or "high_performance".
		411	+
		412	+*@par Outputs:
		413	+*encoded_bboxes: Bboxes generated based on "anchor_boxes" and "gt_bboxes". Have the
		414	+* same format and type as "anchor_boxes".
		415	+*/
		416	+#ifndef OPS_PROTO_DEF_YOLOBOXESENCODE
		417	+#define OPS_PROTO_DEF_YOLOBOXESENCODE
		418	+REG_OP(YoloBoxesEncode)
		419	+ .INPUT(anchor_boxes, TensorType({DT_FLOAT16, DT_FLOAT}))
		420	+ .INPUT(gt_bboxes, TensorType({DT_FLOAT16, DT_FLOAT}))
		421	+ .INPUT(stride, TensorType({DT_INT32}))
		422	+ .ATTR(performance_mode, String, "high_precision")
		423	+ .OUTPUT(encoded_bboxes, TensorType({DT_FLOAT16, DT_FLOAT}))
		424	+ .OP_END_FACTORY_REG(YoloBoxesEncode)
		425	+#endif
		426	+
		427	+/**
		428	+* @brief Generates bounding boxes based on "priors" and "bboxes".
		429	+* It is a customized yolox operator . \n
		430	+
		431	+* @par Inputs:
		432	+* Two inputs, including:
		433	+* @li priors: prior sample boxes of origin image
		434	+* A 2D Tensor of type float32 or float16 with shape (N, 4).
		435	+* @li bboxes_input: bboxes predicted by the model. A 2D Tensor of type float32 or float16 with shape (B, N, 4) . \n
		436	+
		437	+* @par Outputs:
		438	+* bboxes_output: Bboxes generated based on "priors" and "bboxes_input". Have the same format
		439	+* and type as "bboxes_input".
		440	+*/
		441	+#ifndef OPS_PROTO_DEF_YOLOXBOUNDINGBOXDECODE
		442	+#define OPS_PROTO_DEF_YOLOXBOUNDINGBOXDECODE
		443	+REG_OP(YoloxBoundingBoxDecode)
		444	+ .INPUT(priors, TensorType({DT_FLOAT16, DT_FLOAT}))
		445	+ .INPUT(bboxes, TensorType({DT_FLOAT16, DT_FLOAT}))
		446	+ .OUTPUT(decoded_bboxes, TensorType({DT_FLOAT16, DT_FLOAT}))
		447	+ .OP_END_FACTORY_REG(YoloxBoundingBoxDecode)
		448	+#endif
		449	+
		450	+/**
		451	+*@brief Converts ARGB channel . \n
		452	+
		453	+*@par Inputs:
		454	+*@li x: A Tensor. Must be one of the following types: float16.
		455	+
		456	+*@par Outputs:
		457	+*y: A Tensor. Must be one of the following types: int16. \n
		458	+*/
		459	+#ifndef OPS_PROTO_DEF_TRANSARGB
		460	+#define OPS_PROTO_DEF_TRANSARGB
		461	+REG_OP(TransArgb)
		462	+ .INPUT(x, "T1")
		463	+ .OUTPUT(y, "T2")
		464	+ .DATATYPE(T1, TensorType({DT_FLOAT16}))
		465	+ .DATATYPE(T2, TensorType({DT_INT16}))
		466	+ .OP_END_FACTORY_REG(TransArgb)
		467	+#endif
		468	+} // namespace ge
		469	+
230	#endif	470	#endif

Acommon/src/framework/trans_argb_onnx_plugin.cpp+43-0

@@ -0,0 +1,43 @@
	1	+/**
	2	+ * Copyright (c) 2025 Huawei Technologies Co., Ltd.
	3	+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
	4	+ * CANN Open Software License Agreement Version 2.0 (the "License").
	5	+ * Please refer to the License for details. You may not use this file except in compliance with the License.
	6	+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
	7	+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
	8	+ * See LICENSE in the root of the software repository for the full text of the License.
	9	+ */
	10	+
	11	+/*!
	12	+ * \file trans_argb_onnx_plugin.cpp
	13	+ * \brief
	14	+ */
	15	+
	16	+#include "onnx_common.h"
	17	+
	18	+namespace domi {
	19	+using NodeProto = ge::onnx::NodeProto;
	20	+
	21	+static Status ParseParamsTransArgb(const Message* op_src, ge::Operator& op_dest)
	22	+{
	23	+ const NodeProto* node = dynamic_cast<const NodeProto*>(op_src);
	24	+ if (node == nullptr) {
	25	+ OP_LOGE(GetOpName(op_dest).c_str(), "Dynamic cast op_src to NodeProto failed.");
	26	+ return FAILED;
	27	+ }
	28	+
	29	+ return SUCCESS;
	30	+}
	31	+
	32	+// register TransArgb op info to GE
	33	+REGISTER_CUSTOM_OP("TransArgb")
	34	+ .FrameworkType(ONNX)
	35	+ .OriginOpType({ge::AscendString("ai.onnx::8::TransArgb"), ge::AscendString("ai.onnx::9::TransArgb"),
	36	+ ge::AscendString("ai.onnx::10::TransArgb"), ge::AscendString("ai.onnx::11::TransArgb"),
	37	+ ge::AscendString("ai.onnx::12::TransArgb"), ge::AscendString("ai.onnx::13::TransArgb"),
	38	+ ge::AscendString("ai.onnx::14::TransArgb"), ge::AscendString("ai.onnx::15::TransArgb"),
	39	+ ge::AscendString("ai.onnx::16::TransArgb"), ge::AscendString("ai.onnx::17::TransArgb"),
	40	+ ge::AscendString("ai.onnx::18::TransArgb")})
	41	+ .ParseParamsFn(ParseParamsTransArgb)
	42	+ .ImplyType(ImplyType::TVM);
	43	+} // namespace domi

Acommon/src/framework/yolo_boxes_encode_onnx_plugin.cpp+51-0

@@ -0,0 +1,51 @@
	1	+/**
	2	+ * Copyright (c) 2025 Huawei Technologies Co., Ltd.
	3	+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
	4	+ * CANN Open Software License Agreement Version 2.0 (the "License").
	5	+ * Please refer to the License for details. You may not use this file except in compliance with the License.
	6	+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
	7	+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
	8	+ * See LICENSE in the root of the software repository for the full text of the License.
	9	+ */
	10	+
	11	+/*!
	12	+ * \file yolo_boxes_encode_onnx_plugin.cpp
	13	+ * \brief
	14	+ */
	15	+
	16	+#include "onnx_common.h"
	17	+
	18	+namespace domi {
	19	+using NodeProto = ge::onnx::NodeProto;
	20	+
	21	+static Status ParseParamsYoloBoxesEncode(const Message* op_src, ge::Operator& op_dest)
	22	+{
	23	+ const NodeProto* node = dynamic_cast<const NodeProto*>(op_src);
	24	+ if (node == nullptr) {
	25	+ OP_LOGE(GetOpName(op_dest).c_str(), "Dynamic cast op_src to NodeProto failed.");
	26	+ return FAILED;
	27	+ }
	28	+
	29	+ std::string performance_mode = "high_precision";
	30	+ for (const auto& attr : node->attribute()) {
	31	+ if (attr.name() == "performance_mode" && attr.type() == ge::onnx::AttributeProto::STRING) {
	32	+ performance_mode = attr.s();
	33	+ }
	34	+ }
	35	+
	36	+ op_dest.SetAttr("performance_mode", performance_mode);
	37	+ return SUCCESS;
	38	+}
	39	+
	40	+// register YoloBoxesEncode op info to GE
	41	+REGISTER_CUSTOM_OP("YoloBoxesEncode")
	42	+ .FrameworkType(ONNX)
	43	+ .OriginOpType(
	44	+ {ge::AscendString("ai.onnx::11::NPUYoloBoxesEncode"), ge::AscendString("ai.onnx::12::NPUYoloBoxesEncode"),
	45	+ ge::AscendString("ai.onnx::13::NPUYoloBoxesEncode"), ge::AscendString("ai.onnx::14::NPUYoloBoxesEncode"),
	46	+ ge::AscendString("ai.onnx::15::NPUYoloBoxesEncode"), ge::AscendString("ai.onnx::16::NPUYoloBoxesEncode"),
	47	+ ge::AscendString("ai.onnx::17::NPUYoloBoxesEncode"), ge::AscendString("ai.onnx::18::NPUYoloBoxesEncode"),
	48	+ ge::AscendString("npu::1::NPUYoloBoxesEncode")})
	49	+ .ParseParamsFn(ParseParamsYoloBoxesEncode)
	50	+ .ImplyType(ImplyType::TVM);
	51	+} // namespace domi

Acommon/src/framework/yolo_onnx_plugin.cpp+69-0

@@ -0,0 +1,69 @@

1 +/**

3 + * This program is free software, you can redistribute it and/or modify it under the terms and conditions of

4 + * CANN Open Software License Agreement Version 2.0 (the "License").

5 + * Please refer to the License for details. You may not use this file except in compliance with the License.

6 + * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,

7 + * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.

8 + * See LICENSE in the root of the software repository for the full text of the License.

9 + */

10 +

11 +/*!

12 + * \file yolo_onnx_plugin.cpp

13 + * \brief

14 + */

15 +

16 +#include "onnx_common.h"

17 +

18 +namespace domi {

19 +using NodeProto = ge::onnx::NodeProto;

20 +

21 +static Status ParseParamsYolo(const Message* op_src, ge::Operator& op_dest)

22 +{

23 + const NodeProto* node = dynamic_cast<const NodeProto*>(op_src);

24 + if (node == nullptr) {

25 + OP_LOGE(GetOpName(op_dest).c_str(), "Dynamic cast op_src to NodeProto failed.");

26 + return FAILED;

27 + }

28 +

29 + int boxes = 3;

30 + int coords = 4;

31 + int classes = 80;

32 + std::string yolo_version = "V3";

33 + bool softmax = false;

34 + bool background = false;

35 + bool softmaxtree = false;

36 + for (const auto& attr : node->attribute()) {

37 + if (attr.name() == "boxes" && attr.type() == ge::onnx::AttributeProto::INT) {

38 + boxes = attr.i();

39 + } else if (attr.name() == "coords" && attr.type() == ge::onnx::AttributeProto::INT) {

40 + coords = attr.i();

41 + } else if (attr.name() == "classes" && attr.type() == ge::onnx::AttributeProto::INT) {

42 + classes = attr.i();

43 + } else if (attr.name() == "yolo_version" && attr.type() == ge::onnx::AttributeProto::STRING) {

44 + yolo_version = attr.s();

45 + }

46 + }

atomgit-bot7月14日

🟡 Medium Priority

yolo_onnx_plugin.cpp 的 ParseParamsYolo 函数（第36-46行）仅解析了 boxes、coords、classes、yolo_version 四个属性，但 op_cv_proto_extend.h 中 Yolo 算子的 proto 注册（第261-263行）还定义了 softmax、background、softmaxtree 三个 bool 属性。ONNX 模型中若对这些属性设置了非默认值（如 softmax=true），将被静默忽略，导致算子行为与 ONNX 模型语义不一致。

对照同批新增的 yolo_predetection_onnx_plugin.cpp（第36-51行），其 ParseParamsYoloPreDetection 正确解析了全部 7 个属性，进一步表明 yolo_onnx_plugin.cpp 漏解析这三个属性是遗漏而非有意为之。

改动建议

+    for (const auto& attr : node->attribute()) {
+        if (attr.name() == "boxes" && attr.type() == ge::onnx::AttributeProto::INT) {
+            boxes = attr.i();
+        } else if (attr.name() == "coords" && attr.type() == ge::onnx::AttributeProto::INT) {
+            coords = attr.i();
+        } else if (attr.name() == "classes" && attr.type() == ge::onnx::AttributeProto::INT) {
+            classes = attr.i();
+        } else if (attr.name() == "yolo_version" && attr.type() == ge::onnx::AttributeProto::STRING) {
+            yolo_version = attr.s();
+        } else if (attr.name() == "softmax" && attr.type() == ge::onnx::AttributeProto::INT) {
+            softmax = (attr.i() == 1);
+        } else if (attr.name() == "background" && attr.type() == ge::onnx::AttributeProto::INT) {
+            background = (attr.i() == 1);
+        } else if (attr.name() == "softmaxtree" && attr.type() == ge::onnx::AttributeProto::INT) {
+            softmaxtree = (attr.i() == 1);
+        }
     }

应用建议

47 +

48 + op_dest.SetAttr("boxes", boxes);

49 + op_dest.SetAttr("coords", coords);

50 + op_dest.SetAttr("classes", classes);

51 + op_dest.SetAttr("yolo_version", yolo_version);

52 + op_dest.SetAttr("softmax", softmax);

53 + op_dest.SetAttr("background", background);

54 + op_dest.SetAttr("softmaxtree", softmaxtree);

55 +

56 + return SUCCESS;

57 +}

58 +

59 +// register Yolo op info to GE

60 +REGISTER_CUSTOM_OP("Yolo")

61 + .FrameworkType(ONNX)

62 + .OriginOpType({ge::AscendString("ai.onnx::8::Yolo"), ge::AscendString("ai.onnx::9::Yolo"),

63 + ge::AscendString("ai.onnx::10::Yolo"), ge::AscendString("ai.onnx::11::Yolo"),

64 + ge::AscendString("ai.onnx::12::Yolo"), ge::AscendString("ai.onnx::13::Yolo"),

65 + ge::AscendString("ai.onnx::14::Yolo"), ge::AscendString("ai.onnx::15::Yolo"),

66 + ge::AscendString("ai.onnx::16::Yolo")})

67 + .ParseParamsFn(ParseParamsYolo)

68 + .ImplyType(ImplyType::TVM);

69 +} // namespace domi

Acommon/src/framework/yolo_predetection_onnx_plugin.cpp+75-0

@@ -0,0 +1,75 @@
	1	+/**
	2	+ * Copyright (c) 2025 Huawei Technologies Co., Ltd.
	3	+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
	4	+ * CANN Open Software License Agreement Version 2.0 (the "License").
	5	+ * Please refer to the License for details. You may not use this file except in compliance with the License.
	6	+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
	7	+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
	8	+ * See LICENSE in the root of the software repository for the full text of the License.
	9	+ */
	10	+
	11	+/*!
	12	+ * \file yolo_predetection_onnx_plugin.cpp
	13	+ * \brief
	14	+ */
	15	+
	16	+#include "onnx_common.h"
	17	+
	18	+namespace domi {
	19	+using NodeProto = ge::onnx::NodeProto;
	20	+
	21	+static Status ParseParamsYoloPreDetection(const Message* op_src, ge::Operator& op_dest)
	22	+{
	23	+ const NodeProto* node = dynamic_cast<const NodeProto*>(op_src);
	24	+ if (node == nullptr) {
	25	+ OP_LOGE(GetOpName(op_dest).c_str(), "Dynamic cast op_src to NodeProto failed.");
	26	+ return FAILED;
	27	+ }
	28	+
	29	+ int boxes = 3;
	30	+ int coords = 4;
	31	+ int classes = 80;
	32	+ std::string yolo_version = "V5";
	33	+ bool softmax = false;
	34	+ bool background = false;
	35	+ bool softmaxtree = false;
	36	+ for (const auto& attr : node->attribute()) {
	37	+ if (attr.name() == "boxes") {
	38	+ boxes = attr.i();
	39	+ } else if (attr.name() == "coords") {
	40	+ coords = attr.i();
	41	+ } else if (attr.name() == "classes") {
	42	+ classes = attr.i();
	43	+ } else if (attr.name() == "yolo_version") {
	44	+ yolo_version = attr.s();
	45	+ } else if (attr.name() == "softmax") {
	46	+ softmax = attr.i();
	47	+ } else if (attr.name() == "background") {
	48	+ background = attr.i();
	49	+ } else if (attr.name() == "softmaxtree") {
	50	+ softmaxtree = attr.i();
	51	+ }
	52	+ }
	53	+
	54	+ op_dest.SetAttr("boxes", boxes);
	55	+ op_dest.SetAttr("coords", coords);
	56	+ op_dest.SetAttr("classes", classes);
	57	+ op_dest.SetAttr("yolo_version", yolo_version);
	58	+ op_dest.SetAttr("softmax", softmax);
	59	+ op_dest.SetAttr("background", background);
	60	+ op_dest.SetAttr("softmaxtree", softmaxtree);
	61	+
	62	+ return SUCCESS;
	63	+}
	64	+
	65	+// register YoloPreDetection op info to GE
	66	+REGISTER_CUSTOM_OP("YoloPreDetection")
	67	+ .FrameworkType(ONNX)
	68	+ .OriginOpType({ge::AscendString("ai.onnx::8::YoloPreDetection"), ge::AscendString("ai.onnx::9::YoloPreDetection"),
	69	+ ge::AscendString("ai.onnx::10::YoloPreDetection"), ge::AscendString("ai.onnx::11::YoloPreDetection"),
	70	+ ge::AscendString("ai.onnx::12::YoloPreDetection"), ge::AscendString("ai.onnx::13::YoloPreDetection"),
	71	+ ge::AscendString("ai.onnx::14::YoloPreDetection"), ge::AscendString("ai.onnx::15::YoloPreDetection"),
	72	+ ge::AscendString("ai.onnx::16::YoloPreDetection")})
	73	+ .ParseParamsFn(ParseParamsYoloPreDetection)
	74	+ .ImplyType(ImplyType::TVM);
	75	+} // namespace domi

Acommon/src/framework/yolov3_detection_output_v2_onnx_plugin.cpp+108-0

@@ -0,0 +1,108 @@

1 +/**

3 + * This program is free software, you can redistribute it and/or modify it under the terms and conditions of

4 + * CANN Open Software License Agreement Version 2.0 (the "License").

5 + * Please refer to the License for details. You may not use this file except in compliance with the License.

6 + * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,

7 + * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.

8 + * See LICENSE in the root of the software repository for the full text of the License.

9 + */

10 +

11 +/*!

12 + * \file yolov3_detection_output_v2_onnx_plugin.cpp

13 + * \brief

14 + */

15 +

16 +#include "onnx_common.h"

17 +

18 +namespace domi {

19 +using NodeProto = ge::onnx::NodeProto;

20 +

21 +static Status ParseParamsYolov3detectionoutputv2(const Message* op_src, ge::Operator& op_dest)

22 +{

23 + const NodeProto* node = dynamic_cast<const NodeProto*>(op_src);

24 + if (node == nullptr) {

25 + OP_LOGE(GetOpName(op_dest).c_str(), "Dynamic cast op_src to NodeProto failed.");

26 + return FAILED;

27 + }

28 + int n = node->input_size();

29 + op_dest.DynamicInputRegister("x", n);

30 +

31 + int N = 10;

32 + int boxes = 3;

33 + int coords = 4;

34 + int classes = 80;

35 + int post_nms_topn = 512;

36 + int pre_nms_topn = 512;

37 + int out_box_dim = 3;

38 + float obj_threshold = 0.5;

39 + float score_threshold = 0.5;

40 + float iou_threshold = 0.45;

41 + bool relative = true;

42 + bool resize_origin_img_to_net = false;

43 + std::vector<float> v_biases = {};

44 +

45 + for (const auto& attr : node->attribute()) {

46 + if (attr.name() == "boxes" && attr.type() == ge::onnx::AttributeProto::INT) {

47 + boxes = attr.i();

48 + } else if (attr.name() == "coords" && attr.type() == ge::onnx::AttributeProto::INT) {

49 + coords = attr.i();

50 + } else if (attr.name() == "classes" && attr.type() == ge::onnx::AttributeProto::INT) {

51 + classes = attr.i();

52 + } else if (attr.name() == "N" && attr.type() == ge::onnx::AttributeProto::INT) {

53 + N = attr.i();

54 + } else if (attr.name() == "post_nms_topn" && attr.type() == ge::onnx::AttributeProto::INT) {

55 + post_nms_topn = attr.i();

56 + } else if (attr.name() == "pre_nms_topn" && attr.type() == ge::onnx::AttributeProto::INT) {

57 + pre_nms_topn = attr.i();

58 + } else if (attr.name() == "out_box_dim" && attr.type() == ge::onnx::AttributeProto::INT) {

59 + out_box_dim = attr.i();

60 + } else if (attr.name() == "obj_threshold" && attr.type() == ge::onnx::AttributeProto::FLOAT) {

61 + obj_threshold = attr.f();

62 + } else if (attr.name() == "score_threshold" && attr.type() == ge::onnx::AttributeProto::FLOAT) {

63 + score_threshold = attr.f();

64 + } else if (attr.name() == "iou_threshold" && attr.type() == ge::onnx::AttributeProto::FLOAT) {

65 + iou_threshold = attr.f();

66 + } else if (attr.name() == "biases" && attr.type() == ge::onnx::AttributeProto::FLOATS) {

67 + for (auto biases_f : attr.floats()) {

68 + v_biases.push_back(biases_f);

69 + }

70 + }

71 + }

atomgit-bot7月14日

🟡 Medium Priority

yolov3_detection_output_v2_onnx_plugin.cpp 的 ParseParamsYolov3detectionoutputv2 函数（第45-71行）的属性解析循环中缺少对 relative 和 resize_origin_img_to_net 两个 bool 属性的处理。proto 注册（op_cv_proto_extend.h 第387行和第394行）定义了这两个属性（默认值分别为 true 和 false），但插件始终使用本地硬编码的默认值，ONNX 模型中的实际值被静默忽略。

对照同批新增的 yolov5_detection_output_onnx_plugin.cpp（第71-74行），其 ParseParamsYoloV5DetectionOutput 正确解析了 relative 和 resize_origin_img_to_net，进一步表明此处为遗漏。

改动建议

+    for (const auto& attr : node->attribute()) {
+        if (attr.name() == "boxes" && attr.type() == ge::onnx::AttributeProto::INT) {
+            boxes = attr.i();
+        } else if (attr.name() == "coords" && attr.type() == ge::onnx::AttributeProto::INT) {
+            coords = attr.i();
+        } else if (attr.name() == "classes" && attr.type() == ge::onnx::AttributeProto::INT) {
+            classes = attr.i();
+        } else if (attr.name() == "N" && attr.type() == ge::onnx::AttributeProto::INT) {
+            N = attr.i();
+        } else if (attr.name() == "post_nms_topn" && attr.type() == ge::onnx::AttributeProto::INT) {
+            post_nms_topn = attr.i();
+        } else if (attr.name() == "pre_nms_topn" && attr.type() == ge::onnx::AttributeProto::INT) {
+            pre_nms_topn = attr.i();
+        } else if (attr.name() == "out_box_dim" && attr.type() == ge::onnx::AttributeProto::INT) {
+            out_box_dim = attr.i();
+        } else if (attr.name() == "obj_threshold" && attr.type() == ge::onnx::AttributeProto::FLOAT) {
+            obj_threshold = attr.f();
+        } else if (attr.name() == "score_threshold" && attr.type() == ge::onnx::AttributeProto::FLOAT) {
+            score_threshold = attr.f();
+        } else if (attr.name() == "iou_threshold" && attr.type() == ge::onnx::AttributeProto::FLOAT) {
+            iou_threshold = attr.f();
+        } else if (attr.name() == "biases" && attr.type() == ge::onnx::AttributeProto::FLOATS) {
+            for (auto biases_f : attr.floats()) {
+                v_biases.push_back(biases_f);
+            }
+        } else if (attr.name() == "relative" && attr.type() == ge::onnx::AttributeProto::INT) {
+            relative = (attr.i() == 1);
+        } else if (attr.name() == "resize_origin_img_to_net" && attr.type() == ge::onnx::AttributeProto::INT) {
+            resize_origin_img_to_net = (attr.i() == 1);
+        }
     }

应用建议

72 +

73 + if (v_biases.empty()) {

74 + OP_LOGE(GetOpName(op_dest).c_str(), "The attr of biases is required.");

75 + return FAILED;

76 + }

77 +

78 + op_dest.SetAttr("N", N);

79 + op_dest.SetAttr("biases", v_biases);

80 + op_dest.SetAttr("boxes", boxes);

81 + op_dest.SetAttr("coords", coords);

82 + op_dest.SetAttr("classes", classes);

83 + op_dest.SetAttr("relative", relative);

84 + op_dest.SetAttr("post_nms_topn", post_nms_topn);

85 + op_dest.SetAttr("pre_nms_topn", pre_nms_topn);

86 + op_dest.SetAttr("out_box_dim", out_box_dim);

87 + op_dest.SetAttr("obj_threshold", obj_threshold);

88 + op_dest.SetAttr("score_threshold", score_threshold);

89 + op_dest.SetAttr("iou_threshold", iou_threshold);

90 + op_dest.SetAttr("resize_origin_img_to_net", resize_origin_img_to_net);

91 + return SUCCESS;

92 +}

93 +

94 +// register YoloV3DetectionOutputV2 op info to GE

95 +REGISTER_CUSTOM_OP("YoloV3DetectionOutputV2")

96 + .FrameworkType(ONNX)

97 + .OriginOpType({ge::AscendString("ai.onnx::8::YoloV3DetectionOutputV2"),

98 + ge::AscendString("ai.onnx::9::YoloV3DetectionOutputV2"),

99 + ge::AscendString("ai.onnx::10::YoloV3DetectionOutputV2"),

100 + ge::AscendString("ai.onnx::11::YoloV3DetectionOutputV2"),

101 + ge::AscendString("ai.onnx::12::YoloV3DetectionOutputV2"),

102 + ge::AscendString("ai.onnx::13::YoloV3DetectionOutputV2"),

103 + ge::AscendString("ai.onnx::14::YoloV3DetectionOutputV2"),

104 + ge::AscendString("ai.onnx::15::YoloV3DetectionOutputV2"),

105 + ge::AscendString("ai.onnx::16::YoloV3DetectionOutputV2")})

106 + .ParseParamsFn(ParseParamsYolov3detectionoutputv2)

107 + .ImplyType(ImplyType::TVM);

108 +} // namespace domi

Acommon/src/framework/yolov5_detection_output_onnx_plugin.cpp+114-0

@@ -0,0 +1,114 @@
	1	+/**
	2	+ * Copyright (c) 2025 Huawei Technologies Co., Ltd.
	3	+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
	4	+ * CANN Open Software License Agreement Version 2.0 (the "License").
	5	+ * Please refer to the License for details. You may not use this file except in compliance with the License.
	6	+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
	7	+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
	8	+ * See LICENSE in the root of the software repository for the full text of the License.
	9	+ */
	10	+
	11	+/*!
	12	+ * \file yolov5_detection_output_onnx_plugin.cpp
	13	+ * \brief
	14	+ */
	15	+
	16	+#include "onnx_common.h"
	17	+
	18	+namespace domi {
	19	+using NodeProto = ge::onnx::NodeProto;
	20	+
	21	+static Status ParseParamsYoloV5DetectionOutput(const Message* op_src, ge::Operator& op_dest)
	22	+{
	23	+ const NodeProto* node = dynamic_cast<const NodeProto*>(op_src);
	24	+ if (node == nullptr) {
	25	+ OP_LOGE(GetOpName(op_dest).c_str(), "Dynamic cast op_src to NodeProto failed.");
	26	+ return FAILED;
	27	+ }
	28	+ int n = node->input_size();
	29	+ op_dest.DynamicInputRegister("x", n);
	30	+
	31	+ int N = 10;
	32	+ int boxes = 3;
	33	+ int coords = 4;
	34	+ int classes = 80;
	35	+ int post_nms_topn = 512;
	36	+ int pre_nms_topn = 512;
	37	+ int out_box_dim = 3;
	38	+ float obj_threshold = 0.5;
	39	+ float score_threshold = 0.5;
	40	+ float iou_threshold = 0.45;
	41	+ float alpha = 2.0;
	42	+ bool relative = true;
	43	+ bool resize_origin_img_to_net = false;
	44	+ std::vector<float> v_biases = {};
	45	+
	46	+ for (const auto& attr : node->attribute()) {
	47	+ if (attr.name() == "boxes") {
	48	+ boxes = attr.i();
	49	+ } else if (attr.name() == "coords") {
	50	+ coords = attr.i();
	51	+ } else if (attr.name() == "classes") {
	52	+ classes = attr.i();
	53	+ } else if (attr.name() == "N") {
	54	+ N = attr.i();
	55	+ } else if (attr.name() == "post_nms_topn") {
	56	+ post_nms_topn = attr.i();
	57	+ } else if (attr.name() == "pre_nms_topn") {
	58	+ pre_nms_topn = attr.i();
	59	+ } else if (attr.name() == "out_box_dim") {
	60	+ out_box_dim = attr.i();
	61	+ } else if (attr.name() == "obj_threshold") {
	62	+ obj_threshold = attr.f();
	63	+ } else if (attr.name() == "score_threshold") {
	64	+ score_threshold = attr.f();
	65	+ } else if (attr.name() == "iou_threshold") {
	66	+ iou_threshold = attr.f();
	67	+ } else if (attr.name() == "biases") {
	68	+ for (auto biases_f : attr.floats()) {
	69	+ v_biases.push_back(biases_f);
	70	+ }
	71	+ } else if (attr.name() == "relative") {
	72	+ relative = attr.i();
	73	+ } else if (attr.name() == "resize_origin_img_to_net") {
	74	+ resize_origin_img_to_net = attr.i();
	75	+ } else if (attr.name() == "alpha") {
	76	+ alpha = attr.f();
	77	+ }
	78	+ }
	atomgit-bot7月14日 🟡 Medium Priority 文件 `yolov5_detection_output_onnx_plugin.cpp` 中的属性解析循环（第46-78行）在读取所有 14 个属性时均缺少类型检查（即 `attr.type() == ge::onnx::AttributeProto::XXX` 守卫）。这与同一代码库中现有的插件不一致：`yolo_onnx_plugin.cpp`（第37行）和 `yolov3_detection_output_v2_onnx_plugin.cpp`（第46行）对所有属性读取都进行了类型检查。`yolov5_detection_output_onnx_plugin.cpp` 中受影响的属性包括：后果：如果 ONNX 模型包含名称匹配但类型错误的属性（例如，将 `"boxes"` 作为 FLOAT 而非 INT 传递），protobuf 的 `attr.i()` 返回 0（即非整数类型的默认值），`attr.f()` 返回 0.0f。这会导致算子使用错误的值静默运行，输出结果不正确，且无任何错误或警告。对于 `relative`/`resize_origin_img_to_net`，这些是 `bool` 变量，但通过 `attr.i()`（返回 `int64`）读取并隐式转换为 `bool`，如果模型传递了非预期的属性类型，会产生错误结果。
	79	+
	80	+ if (v_biases.empty()) {
	81	+ OP_LOGE(GetOpName(op_dest).c_str(), "The attr of biases is required.");
	82	+ return FAILED;
	83	+ }
	84	+
	85	+ op_dest.SetAttr("N", N);
	86	+ op_dest.SetAttr("biases", v_biases);
	87	+ op_dest.SetAttr("boxes", boxes);
	88	+ op_dest.SetAttr("coords", coords);
	89	+ op_dest.SetAttr("classes", classes);
	90	+ op_dest.SetAttr("relative", relative);
	91	+ op_dest.SetAttr("post_nms_topn", post_nms_topn);
	92	+ op_dest.SetAttr("pre_nms_topn", pre_nms_topn);
	93	+ op_dest.SetAttr("out_box_dim", out_box_dim);
	94	+ op_dest.SetAttr("obj_threshold", obj_threshold);
	95	+ op_dest.SetAttr("score_threshold", score_threshold);
	96	+ op_dest.SetAttr("iou_threshold", iou_threshold);
	97	+ op_dest.SetAttr("resize_origin_img_to_net", resize_origin_img_to_net);
	98	+ op_dest.SetAttr("alpha", alpha);
	99	+
	100	+ return SUCCESS;
	101	+}
	102	+
	103	+// register YoloV5DetectionOutput op info to GE
	104	+REGISTER_CUSTOM_OP("YoloV5DetectionOutput")
	105	+ .FrameworkType(ONNX)
	106	+ .OriginOpType(
	107	+ {ge::AscendString("ai.onnx::8::YoloV5DetectionOutput"), ge::AscendString("ai.onnx::9::YoloV5DetectionOutput"),
	108	+ ge::AscendString("ai.onnx::10::YoloV5DetectionOutput"), ge::AscendString("ai.onnx::11::YoloV5DetectionOutput"),
	109	+ ge::AscendString("ai.onnx::12::YoloV5DetectionOutput"), ge::AscendString("ai.onnx::13::YoloV5DetectionOutput"),
	110	+ ge::AscendString("ai.onnx::14::YoloV5DetectionOutput"), ge::AscendString("ai.onnx::15::YoloV5DetectionOutput"),
	111	+ ge::AscendString("ai.onnx::16::YoloV5DetectionOutput")})
	112	+ .ParseParamsFn(ParseParamsYoloV5DetectionOutput)
	113	+ .ImplyType(ImplyType::TVM);
	114	+} // namespace domi

Acommon/src/framework/yolox_bounding_box_decode_onnx_plugin.cpp+35-0

@@ -0,0 +1,35 @@
	1	+/**
	2	+ * Copyright (c) 2025 Huawei Technologies Co., Ltd.
	3	+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
	4	+ * CANN Open Software License Agreement Version 2.0 (the "License").
	5	+ * Please refer to the License for details. You may not use this file except in compliance with the License.
	6	+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
	7	+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
	8	+ * See LICENSE in the root of the software repository for the full text of the License.
	9	+ */
	10	+
	11	+/*!
	12	+ * \file yolox_bounding_box_decode_onnx_plugin.cpp
	13	+ * \brief
	14	+ */
	15	+
	16	+#include "onnx_common.h"
	17	+
	18	+namespace domi {
	19	+static Status ParseParamsYoloxBoundingBoxDecode(const Message* op_src, ge::Operator& op_dest) { return SUCCESS; }
	atomgit-bot7月14日 🟡 Medium Priority `ParseParamsYoloxBoundingBoxDecode` 函数（第19行）直接返回 `SUCCESS`，未通过 `dynamic_cast<const NodeProto>(op_src)` 验证 `op_src` 参数，也未进行空指针检查。这与代码库中所有其他 ONNX 插件不一致，包括那些无属性需要解析的插件。例如，`trans_argb_onnx_plugin.cpp`（第21-27行）虽然没有任何属性需要提取，但仍然执行了 `dynamic_cast` 和空指针检查。`affine_grid_onnx_plugin.cpp` 也遵循此模式。后果：如果 GE 框架向 `ParseParamsFn` 回调传递了损坏的或非 `NodeProto` 消息，该函数会静默返回 `SUCCESS`，而其他所有插件都会捕获此错误并记录日志。如果该算子后续增加了属性，这种缺失的验证将成为一个潜在缺陷——未来的维护者可能不会意识到需要添加此检查，因为函数签名已经接受了 `op_src`，而其内容却从未被验证。建议：在 `ParseParamsYoloxBoundingBoxDecode` 内添加 `dynamic_cast` 和空指针检查，与所有其他插件保持一致。具体为：将函数体替换为与 `ParseParamsTransArgb` 等效的三行代码：对 `op_src` 执行 `dynamic_cast<const NodeProto>(op_src)`、空指针检查并记录日志、最后返回 `SUCCESS`。
	20	+
	21	+// register YoloxBoundingBoxDecode op info to GE
	22	+REGISTER_CUSTOM_OP("YoloxBoundingBoxDecode")
	23	+ .FrameworkType(ONNX)
	24	+ .OriginOpType({ge::AscendString("ai.onnx::8::YoloxBoundingBoxDecode"),
	25	+ ge::AscendString("ai.onnx::9::YoloxBoundingBoxDecode"),
	26	+ ge::AscendString("ai.onnx::10::YoloxBoundingBoxDecode"),
	27	+ ge::AscendString("ai.onnx::11::YoloxBoundingBoxDecode"),
	28	+ ge::AscendString("ai.onnx::12::YoloxBoundingBoxDecode"),
	29	+ ge::AscendString("ai.onnx::13::YoloxBoundingBoxDecode"),
	30	+ ge::AscendString("ai.onnx::14::YoloxBoundingBoxDecode"),
	31	+ ge::AscendString("ai.onnx::15::YoloxBoundingBoxDecode"),
	32	+ ge::AscendString("ai.onnx::16::YoloxBoundingBoxDecode")})
	33	+ .ParseParamsFn(ParseParamsYoloxBoundingBoxDecode)
	34	+ .ImplyType(ImplyType::TVM);
	35	+} // namespace domi