Android AOSP Tensorflow on-device training

[Ganesh Ramani]

Hi,

Android AOSP R release I found 2.1.0 support of Tensorflow, 

"https://android.googlesource.com/platform/external/tensorflow/+/refs/heads/master/tensorflow/core/public/version.h"

I have some basic questions :

1. Is the library Tensorflow or Tensorflow-lite ?

2. Will this library supports on device training ? 

3. If any on-device training disabled can I enable and further can I use on my device ?

Thanks,

Ganesh

[Yu-Cheng Ling]

Hi Ganesh,

Thanks for reaching out. 

It would be helpful to understand a bit more about your use case to comment more accurately. 

(E.g. Are you developing an Android app? Do you want to train models on devices? and what kind of models?) 

On Mon, Feb 1, 2021 at 6:42 AM Ganesh Ramani <gramani....@gmail.com> wrote:

Hi,

Android AOSP R release I found 2.1.0 support of Tensorflow, 

"https://android.googlesource.com/platform/external/tensorflow/+/refs/heads/master/tensorflow/core/public/version.h"

I have some basic questions :

1. Is the library Tensorflow or Tensorflow-lite ?

AOSP maintains a fork of TensorFlow (including TensorFlow LIte) codebase for AOSP's internal usage. 

There's no easy way to access this library from Android apps, and it doesn't help for on-device training. 

 

2. Will this library supports on device training ? 

3. If any on-device training disabled can I enable and further can I use on my device ?

We are working on supporting on-device training in TensorFlow Lite, and we will post updates in the TensorFlow Lite discussion group as we make progress. 

Meanwhile, you can take a look at Example on-device model personalization with TensorFlow Lite if you're interested. 

Best,

YC

 

Thanks,

Ganesh

--
You received this message because you are subscribed to the Google Groups "TensorFlow Lite" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tflite+un...@tensorflow.org.
To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/tflite/e70331f1-4e48-4428-9093-06e6bb3ef78an%40tensorflow.org.

[Ganesh Ramani]

Hi Yu-Cheng,

Thanks for your reply,

Here my brief "Use-case" detail:

Have data around 100000 x 20, and want to predict some pattern on my device (kind of in vehicle infotainment h/w) , the data will be trained once in a month or after a certain number of records accumulation. This predicted NN model will be running as a service, Android application (using AIDL) invokes the service to trigger prediction, trigger training.

For the use case, I want to use the Tensorflow library with support of on-device training. I got the information already tensorflow-lite does not support as of now.

Could you please confirm for my use case the current Tensorflow library (2.1 which in AOSP) does not support on-device training or any other short of training ?

Thanks,

Ganesh

[Ganesh Ramani]

Yu-Cheng,

Along with my previous questions ! have one more,

 Can I use the Tensorflow library (TF source) for on-device training ? (assume I've only with CPU) If yes, I can use the forked Tensorflow source of AOSP as well, right ?

Thanks,

Ganesh

[Yu-Cheng Ling]

Hi Ganesh, 

Technically it's possible to build and run TensorFlow on Android devices and perform on-device training. 

However please be advised that this is not officially supported by TensorFlow, and there is no easy to use APIs, so it might be very difficult to make this work. 

As mentioned earlier, we're working hard to enable on-device training in TensorFlow Lite, and hopefully there will be a better way to train on devices soon. 

Thanks,

YC

[Ganesh Ramani]

  HI Yu-Cheng,

Thanks for your reply and advice,

Could you please tell me what are the challenges here if someone takes to make work Tensorflow on device. Any link to refer ?

Kindly share me,  

Thanks,

Ganesh

[Ganesh Ramani]

HI Yu-Cheng,

Thanks for your reply and advice,

Could you please tell me what are the challenges here if someone takes to make Tensorflow work ondevice. Any link to refer ?

Kindly share me,  

Thanks,

Ganesh