topic Re: I.MX8M plus AI Robot platform tensor flow lite NNAPI accelerate issue in i.MX Processors

I.MX8M plus AI Robot platform tensor flow lite NNAPI accelerate issue

Arrow_AE_KingLiu — Fri, 11 Jun 2021 03:25:43 GMT

Dear xiaodong

I compile the 5.4.70 AI robot ROS image using the command "bitbake imx-robot-sdk" ,and I update the OS image with UUU tool, i can brinig up the AI robot EVK successully .

Then I use the command "bitbake imx-robot-system -c populate_sdk" to generate the SDK , i have installed the SDK in my X86 ubuntu 18.04 .

Then I download the tensorflow lite v2.3.1 source code :

git clone https://github.com/tensorflow/tensorflow.git

git checkout v2.3.1

cd /4Tdisk/king/NXP/King_application/tensorflow/tensorflow/lite/examples/label_image/

Compiled the label_image successfully using the command below :

$CC label_image.cc bitmap_helpers.cc ../../tools/evaluation/utils.cc -I=/usr/include/tensorflow/lite/tools/make/downloads/flatbuffers/include \

-I=/usr/include/tensorflow/lite/tools/make/downloads/absl -O1 -DTFLITE_WITHOUT_XNNPACK -ltensorflow-lite -lstdc++ -lpthread -lm -ldl -lrt

After i get the label_image output application a.out , I copy it to the AI robot EVK .

It can run successufly using the CPU .

But can not use the NNAPI accelerate .

But the Yocot pre-build label_image can use the NNAPI accelerate .

How can i debug this issue?

Re: I.MX8M plus AI Robot platform tensor flow lite NNAPI accelerate issue

xiaodong_zhang — Fri, 11 Jun 2021 05:23:01 GMT

@Arrow_AE_KingLiu

Current i.MX Yocto don't support Tensorflow with NPU accelerate(I don't think it will be supported in future). You need to add tensorflow-lite into variable IMAGE_INSTALL in imx-robot-sdk.bb when you want to deploy model trained and converted by TensorFlow tools.

BTW, I already add ML packages (include tensorflow-lite) in imx-robot-sdk.bb in last i.MX robot platform.

https://community.nxp.com/t5/i-MX-Processors-Knowledge-Base/i-MX-%E6%9C%BA%E5%99%A8%E4%BA%BA%E5%B9%B3%E5%8F%B0-i-MX-Robot-Platform/ta-p/1198619

Please the below update

/////////////////////////////////////// update for v1.0-L5.4.70-2.3.2 ///////////////////////////////////////////////////////////

Re: I.MX8M plus AI Robot platform tensor flow lite NNAPI accelerate issue

xiaodong_zhang — Fri, 11 Jun 2021 05:48:25 GMT

@Arrow_AE_KingLiu

Maybe you miss -DTFLITE_ENABLE_NNAPI=on

Re: I.MX8M plus AI Robot platform tensor flow lite NNAPI accelerate issue

Bio_TICFSL — Fri, 11 Jun 2021 13:16:38 GMT

Good

Re: I.MX8M plus AI Robot platform tensor flow lite NNAPI accelerate issue

Arrow_AE_KingLiu — Fri, 18 Jun 2021 07:54:10 GMT

@xiaodong_zhang

We must use the label_image source code located in the yocto bitbake path , then it can use the NNAPI acceleration .

For the label_image source code download from tensorflow lite website , it is different from the one in Yocto BSP .

Re: I.MX8M plus AI Robot platform tensor flow lite NNAPI accelerate issue

bret_harte — Mon, 04 Apr 2022 06:12:58 GMT

Can you be more specific where I can find the version of label_image source code that works with NNAPI acceleration, thanks.