Infer wrong values of bounding box on NPU

取消
显示结果 
显示  仅  | 搜索替代 
您的意思是: 
已解决

Infer wrong values of bounding box on NPU

跳至解决方案
2,007 次查看
MiloJan
Contributor I

Hello NXP,

Our program can use the NPU to achieve acceleration using the tflite VX delegate, but can't obtaine the same result as the CPU.

The label value can be obtained correctly, but the value of the bounding box is always 0.

Our environment:

Board: i.MX 8M Plus EVK

System: kirkstone-5.15.71-2.2.0

Tflite Mode: Yolov v7.0

Is there something we didn't notice?

0 项奖励
回复
1 解答
1,829 次查看
Dhruvit
NXP TechSupport
NXP TechSupport

Hi @MiloJan,

I hope you are doing well.

Please find a response from the engineering team below.

One can use the attached patch to fix the issue on the 5.15.71_2.2.0. Use the Yocto SDK to build the libvx_delegate.so.

The steps are as below:

1. Setup the Yocto SDK 5.15.72.(Refer to the i.MX Yocto Project User's Guide)

2. bitbake the vx_delegate, and use the patch by the command

bitbake tensorflow-lite-vx-delegate

cd imx-yocto-bsp/8mp_5.15.72_2.2.0/build-xwayland/tmp/work/armv8a-poky-linux/tensorflow-lite-vx-delegate/2.9.1-r0/git

git apply diff.patch

bitbake -c compile -f tensorflow-lite-vx-delegate

bitbake -c build -f tensorflow-lite-vx-delegate

the libvx_delegate.so will be generated in ../image/usr/lib/

3.copy the libvx_delegate to the broad path /usr/lib/

Thanks & Regards,
Dhruvit Vasavada

在原帖中查看解决方案

0 项奖励
回复
5 回复数
1,909 次查看
MiloJan
Contributor I

Hello @Dhruvit ,

How is it. Could you reproduce that?

0 项奖励
回复
1,830 次查看
Dhruvit
NXP TechSupport
NXP TechSupport

Hi @MiloJan,

I hope you are doing well.

Please find a response from the engineering team below.

One can use the attached patch to fix the issue on the 5.15.71_2.2.0. Use the Yocto SDK to build the libvx_delegate.so.

The steps are as below:

1. Setup the Yocto SDK 5.15.72.(Refer to the i.MX Yocto Project User's Guide)

2. bitbake the vx_delegate, and use the patch by the command

bitbake tensorflow-lite-vx-delegate

cd imx-yocto-bsp/8mp_5.15.72_2.2.0/build-xwayland/tmp/work/armv8a-poky-linux/tensorflow-lite-vx-delegate/2.9.1-r0/git

git apply diff.patch

bitbake -c compile -f tensorflow-lite-vx-delegate

bitbake -c build -f tensorflow-lite-vx-delegate

the libvx_delegate.so will be generated in ../image/usr/lib/

3.copy the libvx_delegate to the broad path /usr/lib/

Thanks & Regards,
Dhruvit Vasavada

0 项奖励
回复
1,784 次查看
MiloJan
Contributor I

Hello @Dhruvit,

This works for us. Thanks.

 

 

0 项奖励
回复
1,878 次查看
Dhruvit
NXP TechSupport
NXP TechSupport

Hi @MiloJan,

I hope you are doing well.

I have reproduced it and I think the root cause is the vxdelegate driver update.

I am checking the two branch differences and the commit log.

I will update here ASAP.

Thanks & Regards,
Dhruvit Vasavada

0 项奖励
回复
1,982 次查看
Dhruvit
NXP TechSupport
NXP TechSupport

Hi @MiloJan,

I hope you are doing well.

I am trying to reproduce this issue on my end.

I will get back to you once done.

Thanks & Regards,
Dhruvit Vasavada

 

0 项奖励
回复