[Bug-Report|缺陷反馈]: BatchMultiClassNonMaxSuppression 与 SortedNMS InferShape 旧协议兼容问题-ops-cv-AtomGit

已关闭

[Bug-Report|缺陷反馈]: BatchMultiClassNonMaxSuppression 与 SortedNMS InferShape 旧协议兼容问题 #888

黄晓彬创建于 12 天前关闭于 12 天前

12 天前创建

Thanks for sending an issue! Please fill in the following template to help quickly solve your problem.

Describe the current behavior / 问题描述 (Mandatory / 必填)

BatchMultiClassNonMaxSuppression 新 InferShape 固定输出 boxes [B,M,4]、count [B]，与旧 GE 融合内部 transpose_box=true 所需的 boxes [B,4,M]、count [B,8] 不一致。950 实现的 true 分支还需同步按 scores [B,C,N] 解析和访问数据。

此外，Batch 对输入 rank、batch 的强校验，以及 SortedNMS 对 rank 和 ShapeRange 已知上界的强校验，可能拒绝旧通路可继续推导的动态输入。本 Issue 跟踪这些推导与旧有效协议的兼容问题，不改变 NMS 筛选算法。

Environment / 环境信息 (Mandatory / 必填)

问题反馈：A2 上的 BatchMultiClassNonMaxSuppression用例执行失败
版本：CANN 最新周版本

Steps to reproduce the issue / 重现步骤 (Mandatory / 必填)

1 执行 BatchMultiClassNonMaxSuppression 用例

import torch
import torch_npu
import numpy as np
a = np.fromfile(
    "batchnms_L0_xlsx_fp16_1_100_50_4_1_100_50_fp16_0d0_2d0_0d0_1d0_u_ND_0d05_0d5_300_200_input_0.bin",
    np.float16).reshape(1, 100, 50, 4)
b = np.fromfile("batchnms_L0_xlsx_fp16_1_100_50_4_1_100_50_fp16_0d0_2d0_0d0_1d0_u_ND_0d05_0d5_300_200_input_1.bin",
    np.float16).reshape(1,100,50)
score_threshold = 0.05
iou_threshold = 0.5
max_size_per_class = 300
max_total_size = 200
a_npu = torch.from_numpy(a).npu()
b_npu = torch.from_numpy(b).npu()
output_data1, output_data2, output_data3, output_data4 = torch_npu.npu_batch_nms(a_npu, b_npu,
                                                                                 score_threshold,
                                                                                     iou_threshold,
                                                                                     max_size_per_class,
                                                                                     max_total_size)

Describe the expected behavior / 预期结果 (Mandatory / 必填)

BatchMultiClassNonMaxSuppression 用例在A2上面跑通

[ERROR] APP(2780633,python3):2026-09-02-11:20:25.150.213 2780795 OpParamMaker.cpp:printErrorLog:442: "[PTA]:---OpName---BatchMultiClassNonMaxSuppression"
[ERROR] APP(2780633,python3):2026-09-02-11:20:25.150.244 2780795 OpParamMaker.cpp:printErrorLog:460: "[PTA]:InputDesc[0]: DescType = ACL_FLOAT16, DescFormat = ACL_FORMAT_NCHW, DescShape = [1, 100, 50, 4]"
[ERROR] APP(2780633,python3):2026-09-02-11:20:25.150.249 2780795 OpParamMaker.cpp:printErrorLog:460: "[PTA]:InputDesc[1]: DescType = ACL_FLOAT16, DescFormat = ACL_FORMAT_ND, DescShape = [1, 100, 50]"
[DEBUG] APP(2780633,python3):2026-09-02-11:20:25.150.256 2780795 OpParamMaker.cpp:ExecFunc:568: "[PTA]:ExecFunc: Op BatchMultiClassNonMaxSuppression Run with aclopCompileAndExecute, ret = 500002."
[ERROR] APP(2780633,python3):2026-09-02-11:20:25.150.279 2780795 NPUQueue.cpp:ReadQueue:459: "[PTA]:---Thread---281470857703808: device = 0, write_idx = 1, read_idx = 0, status = 0, ret = 500002"
[INFO] APP(2780633,python3):2026-09-02-11:20:25.150.283 2780795 NPUQueue.cpp:ReadQueue:461: "[PTA]:ReadQueue: read failed, the current working operator name is BatchMultiClassNonMaxSuppression, device = 0, write_idx = 1, read_idx