Thanks for sending an issue! Please fill in the following template to help quickly solve your problem.
The Conv3DV2 operator in the operator repository exhibits a significant accuracy problem when processing specific input configurations.
NA
Use Torch NPU to run the case as below
inputs [7, 15, 8, 8, 9] filters [17, 15, 7, 1, 6] bias [17]
The Conv3DV2 operator should produce numerical results with acceptable precision that match the expected outputs from a reference implementation
Using the precision tool to test, got the result as below
'error_info': 'ERROR: 最大相对误差比例(2400344.25)超过阈值(2.0)，\nERROR: 平均相对误差比例(25159.982421875)超过阈值(1.2)，\nERROR: 均方根误差比例(2049177.0)超过阈值(1.2)
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)
The Conv3DV2 operator in the operator repository exhibits a significant accuracy problem when processing specific input configurations.
Environment / 环境信息 (Mandatory / 必填)
NA
Steps to reproduce the issue / 重现步骤 (Mandatory / 必填)
Use Torch NPU to run the case as below
Describe the expected behavior / 预期结果 (Mandatory / 必填)
The Conv3DV2 operator should produce numerical results with acceptable precision that match the expected outputs from a reference implementation
Related log / screenshot / 日志 / 截图 (Mandatory / 必填)
Using the precision tool to test, got the result as below
'error_info': 'ERROR: 最大相对误差比例(2400344.25)超过阈值(2.0)，\nERROR: 平均相对误差比例(25159.982421875)超过阈值(1.2)，\nERROR: 均方根误差比例(2049177.0)超过阈值(1.2)
Special notes for this issue/备注 (Optional / 选填)