Tensorflow example not working on nRF Connect SDK

[kclauah]

I have tried tensorflow example on nRF Connect SDK v1.6.1, v1.7.0 and v1.7.1 but all of them are not working.

To reproduce the behavior, build any example under zephyr\sample\modules\tensorflow

Here is the error from output console:

1> Compiling ‘model.cc’
2> Compiling ‘output_handler.cc’
4> ../src/main_functions.cc:19:10: fatal error: tensorflow/lite/micro/all_ops_resolver.h: No such file or directory
4> compilation terminated.
2> In file included from ../src/output_handler.cc:17:
2> ../src/output_handler.h:20:10: fatal error: tensorflow/lite/c/common.h: No such file or directory
2> compilation terminated.
Build failed

[Simon]

I can see that CONFIG_TENSORFLOW_LITE_MICRO will get the value 'n', even if it's set to 'y'. I'm not sure why this is. Maybe that's what's leading to the issues you see.

However, since this is a Zephyr sample, I think you'll get better help if you create a Zephyr issue or ask in any of these resources: https://developer.nordicsemi.com/nRF_Connect_SDK/doc/latest/zephyr/introduction/index.html#resources.

Check out this ticket as well: https://devzone.nordicsemi.com/f/nordic-q-a/79892/tensorflow-lite-micro-support-for-machine-learning/330840#330840 

Best regards,

Simon

[kclauah]

Thank you for the suggestions.

I have created an issue in zephyr github.

https://github.com/zephyrproject-rtos/zephyr/issues/40827

But the error

CONFIG_TENSORFLOW_LITE_MICRO was assigned the value y, but got the value n. Missing dependencies: 0

does it mean the tensorflow lite module is missing? I have tried searching all_ops_resolver.h but I got nothing. Maybe I can download the whole repo from https://github.com/tensorflow/tensorflow and put it in somewhere, is it correct?

[kclauah]

Thank you for the response.

I have followed the getting started guide instructions and install the Zephyr upstream using Ubuntu 20.0.4.

The tensorflow hello_world sample can be compiled correctly using nrf52dk_nrf52832 and nrf9160dk_nrf9160ns.

I have also tried downloading the module from https://github.com/zephyrproject-rtos/tensorflow and put it in <root>\modules\lib\tensorflow but still showing the same error.

Any suggestions is appreciated.

[kclauah]

The tensorflow example works after following the suggestions from @laurenmurphyx64. It is a bug from nrf-sdk and still exists in main stream.

See the GitHub issue for the solutions.

github.com/.../40827

@Simon How can I report a bug here?

[Simon]

Yes, of course, that would be the solution, adding TensorFlow to name-allowlist. I'm sorry for not informing you about this earlier. I tried to add TensorFlow to https://github.com/nrfconnect/sdk-nrf/blob/v1.8.0-rc1/west.yml#L18 and https://github.com/nrfconnect/sdk-nrf/blob/v1.8.0-rc1/west.yml#L96, which I wasn't able to get to work for some reason.

I do not think this is a bug, but rather an active decision by our developers to not include it in the name-allowlist. We do have better support for Edge Impulse samples in the nRF Connect SDK. Check out our Edge Impulse samples here: https://developer.nordicsemi.com/nRF_Connect_SDK/doc/latest/nrf/samples/edge_impulse/data_forwarder/README.html 

Best regards,

Simon

[kclauah]

So what is the reason of not including tensorflow in the name-allowlist? Edge Impulse is also a machine learning framework but why it does not need the tensorflow lite framework?

[Simon]

I talked to a developer about this. He informed me that the Edge Impulse (EI) actually uses TensorFlow and it's even stated on their main page (https://www.edgeimpulse.com/)

The EON compiler ued by Edge Impulse is even better than TensorFlow Lite (it's more optimized for embedded systems).

Also, you could use TensorFlow's Keras APIs to edit the ML model in EI studio.

If you want to learn more, take a look at this: https://blog.tensorflow.org/2021/06/how-tensorflow-helps-edge-impulse-make-ml-accessible.html

[kclauah]

Sorry. I am not familiar with Edge Impulse.

I misunderstand because the Edge Impulse example KConfig does not include tensorflow lite library. And it can be successfully build and run. So I assume they have their own implementation of ML framework.

Even EON compiler is more optimized than tensorflow lite, I do not think it is reasonable to disable tensorflow. At least, developer has an option to use it or not, right?

Thank you for your patience.