This is a cache of https://discuss.96boards.org/t/hikey970-ddk-support-issue/8008. It is a snapshot of the page at 2024-11-01T02:13:59.500+0000.
Hikey970 DDK support issue - HiKey970 - 96Boards Forum

Hikey970 DDK support issue

Products Support HiKey970
#1

Board based on the Kirin 970 - HI3670 Application Processor
I try to follow the steps in https://developer.huawei.com/consumer/cn/devservice/doc/3140204#2.4 to do the OperatorsCheck with tensorflow tools, but get some conflicts.
No matter what .pb models, from /tools_tensorflow(in that DDK tool file) or from the TFLite hosted model(Hosted models  |  TensorFlow Lite)
Even from the example offered by the HiAI DDK(same link as first one and in section 4.3.1) cannot pass that check. [THIS RESULT IS DIFFERENT FROM THE OFFICIAL WEBSITE!!]

123
123971×451 11.1 KB

Does someone try to fix this error or get the point to design the model that can pass this check?
I’ve tried that in Win10, ubuntu Linux 16.04, macOS but fail.

#2

Hi @11119

In the website you referred, they have stated the requirement of NPU enabled device for running the tensorflow model. So far none of the debian/ubuntu builds for Hikey970 has NPU enabled and that could be the reason for the failures you encountered.

On the other hand we do have prebuilt tensorflow based images optimized for CPU which you can try.

http://www.lemaker.org/product-hikey970-download-86.html

1 Like
#3

Thanks for the reply.
I am using the Hikey970 + AOSP (android 8.1 up) so I want to try this method.

By the way, I’ve use the debian with prebuilt tensorflow. Is there anyway to specify the NPU to run the TensorFlow?

#4

I’ve tried that in Win10, ubuntu Linux 16.04, macOS but fail
=> this means I used these as host system and want to build the APP for the Hikey970 with AOSP to run it.
I’m so sorry i didn’t make myself clear. Please don’t be mistaken.

#5

Tensorflow with the debian image can use the GPU (using OpenCL) but no NPU.
NPU has no support under Linux.

I have not tested under AOSP. Perhaps if you get some results you might want to share with us?

#6

Sure. Thanks for the sharing of the using in Linux!

I also ask the question to Huawei but didn’t receive the reply.
I’m not sure how long it take to fix the problem.
HAHA!

If I can understand Huawei’s tricks, I could share the experience.