Run Yolox-nano on khadas VIM3

VIM3
#1

Hello,

I am trying to run my yolox-nano model on khadas vim3 using npu,
but I am having a hard time trying to do that,

I am using Tengine and TIM-VX to do that (this is the only way I found !!),
if there is a better way please guide me,

any help is appreciated,
thank you!.

#2

@Mohammad_Nour_Alawad

Please follow those docs.

#3

hello @Frank ,

I have been trying to run the aml_npu_demo_binaries and I faced some problems,

the steps I followed are:

  • Install this image to sd card for booting:
    VIM3_Ubuntu-gnome-focal_Linux-4.9_arm64_SD-USB_V1.0.10-220108.img

  • Clone the repository.

  • Run (sudo ./INSTALL) in the detect_demo_picture

  • Trying this command:
    ./detect_demo_x11 0 1080p.bmp

and here is the output …

image
image1481×543 120 KB

what could be the problem?

thanks.

#4

@Mohammad_Nour_Alawad Check the Galcore version

$ dmesg | grep Galcore

Do you upgrade your system to least OTA version?

#5

hello @Frank ,

I noticed that there is no demo for YoloX in the aml_npu_demo_binaries nor in the aml_npu_app where I should do the compilation,

so, what should I do?,
is there something I am missing?

so far, I managed to run the demo binaries on my vim3 pro,
and I built a YoloX-nano model,
now, how can I run this model on the vim3 pro?

thank you.

#6

@Mohammad_Nour_Alawad

You can follow this doc . But the postpress just for yolov3 , if you use another model , please implement your own post-processor.

#7

hello @Frank ,

I am trying to run a yolov3 model on my khadas vim3 pro USING NPU,

I followed this step by step:

and after the compilation is done with no errors,

  • I converted my model to tmfile,
  • I quantized the model to uint8
  • I tried to run ./build/install/bin/tm_yolov3_uint8 and it WORKED perfectly.

but when I try to run ./build/install/bin/tm_yolov3_timvx I am getting this error:

tengine-lite library version: 1.5-dev
E [query_hardware_caps:50]CHECK STATUS(-10:The supplied parameter information does not match the kernel contract.)
E [Init:194]Create tensor fail!
E [Init:194]Create tensor fail!
E [Init:194]Create tensor fail!

P.S:
when I did the build of Tengine I sat:

  • cmake -DTENGINE_ENABLE_TIM_VX=ON

and I checked the Galcore version and It is the latest 6.4.8.7.1.1.1

could you please tell what could be the problem?

thanks a lot!

#8

@Mohammad_Nour_Alawad Please use xxx_uint8 code, other code not for VIM3.

#9

Actually that is very shocking,

It is literally mentioned in the documentation that if you want the NPU to be used then you should use the TIM-VX plugin,

Is the code xxxx_uint8 using the NPU? and how can I make sure?

#10

@Mohammad_Nour_Alawad You can check it source code. If it is a NPU demo, you can found some configiure with NPU.