C library for Android about hello_tf_c_api

[Neargye]

TFLite is a library more suitable for mobile devices. Maybe you try it for android?
I use TF on macos/win, and TFLite on ios/android.

[dtczhl]

Does TFLite use the same API as TF C?
I'm working on a project with a headset that runs Android OS. The SDK of that headset allows me to link to .so libraries. That's why I like to know if there is a similar libtensorflow.so for Android.
If it is convenient for you, could you kindly add TFLite usage to your repo?
It'll help me a lot, and I think other people would also be interested in how to use TFLite in Android.

[Neargye]

You need use tf in c++? Or in java?

[dtczhl]

In C/C++

[Neargye]

I try add my example.
For now you can try this http://www.sanj.ai/Android-App-With-Tflite-C++-API/

[dtczhl]

Thanks very much. Could you let me know after you add your examples?
Your tutorials are very concise and clear.

[Neargye]

Yes, I update this issues when adding android example

[dtczhl]

Yes, I update this issues when adding android example

Hi. I'm wondering the progress of updating Tensorflow Lite. No rush though.

BTW, your documents (doc/optimizing.md and doc/prepare_models.md) on mobile devices are detailed and very helpful. I appreciate that.

[dtczhl]

Yes, I update this issues when adding android example

Hi. I can run TF Lite C++ in Android now, but your tutorials are still welcome.
BTW, are you able to run batch inputs in TF Lite?
For example, a model takes an input-length of 10. I can use Tensorflow C to feed 50 numbers, which runs 5 times of the model with a single function call (which is much faster than manually running the model 5 times).
I see TFLite Interpreter has a function called ResizeInputTensor, which I think is for this purpose. But I got errors when I try to input more numbers. Is there anything that might easily go wrong for beginners?

[Neargye]

Hi, check this batch example https://github.com/Neargye/hello_tf_c_api/blob/master/src/batch_interface.cpp

[dtczhl]

Hi, check this batch example https://github.com/Neargye/hello_tf_c_api/blob/master/src/batch_interface.cpp

It works, but it is for Tensorflow C. Do you have examples for TF Lite?
Thanks.

[Neargye]

Try something like https://stackoverflow.com/questions/51576944/tensorflow-lite-for-variable-sized-input

[dtczhl]

Try something like https://stackoverflow.com/questions/51576944/tensorflow-lite-for-variable-sized-input

I tried that before. However, my program crashes whenever I call

ResizeInputTensor();
AllocateTensors();

even I use the exact dimensions without using batches.
For example, my model takes input of dimensions 1, 100, 9
I tried to run

this->interpreter->ResizeInputTensor(this->interpreter->inputs()[0], {1, 100, 9});
this->interpreter->AllocateTensors();

it crashes. However, if I just call

this->interpreter->AllocateTensors();

then everything works fine. Not sure what goes wrong.

[dtczhl]

I changed Tensorflow Lite version to 2.0, which allows me to run

this->interpreter->ResizeInputTensor(this->interpreter->inputs()[0], {1, 100, 9});
this->interpreter->AllocateTensors();

Basically, nothing is changed because I just call ResizeInputTensor with original/default dimensinos.

However, I cannot change to other dimensions. This is still an open issue here
tensorflow/tensorflow#22377
It seems that Android Tensorflow Lite does not support it.

[Neargye]

could you please send project example, I will try myself

[dtczhl]

could you please send project example, I will try myself

I attach my sample project below. Thanks.
sample project

[Neargye]

Hi, I do some test, it seems you are right, Android Tensorflow Lite does not support it.

[dtczhl]

Hi, I do some test, it seems you are right, Android Tensorflow Lite does not support it.

Thanks. BTW, do you know if it works in iOS? I'm just curious about that.

[Neargye]

Hm, I can check tomorrow in iOS.

[dtczhl]

Hm, I can check tomorrow in iOS.

Thank you. Please let me know after you check it.

[Neargye]

I could not do it on iOS.

[dtczhl]

I could not do it on iOS.

Thanks.