TFLite Model personalization on IoT devices

312 views
Skip to first unread message

Ivelin Ivanov

unread,
Mar 23, 2021, 1:03:46 PM3/23/21
to tfl...@tensorflow.org
Hello tflite team!

Thank you for building an awesome project.

Do you have roadmap visibility for model personalization (transfer learning) on IoT devices (e.g. rpi) ? The feature has been available for Android since 2019 and the blog post hints to an upcoming implementation for other platforms.

Future work
The evolution of the transfer learning pipeline in the future is likely to be coupled with the development of the full training solution in TensorFlow Lite. Today we provide the transfer learning pipeline as a separate example on GitHub, and in the future we plan to support full training. The transfer learning converter would then be adapted to produce a single TensorFlow Lite model that would be able to run without an additional runtime library.

Thank you,

Ivelin

Alex

unread,
Mar 23, 2021, 6:45:30 PM3/23/21
to TensorFlow Lite, ivelin...@ambianic.ai

Ivelin Ivanov

unread,
Mar 26, 2021, 6:05:30 PM3/26/21
to Alex, TensorFlow Lite
Thank you, Alex. I was asking how to use tflite for transfer learning, without installing the full tf/keras package.

The most up to date answer I found was provided by the TFLite Support team:

TFLite Model Maker is now available for IoT devices (e.g. rpi) with ability to do transfer learning on three tasks: image classification, text classification and Q&A.

Ivelin Ivanov

unread,
Mar 26, 2021, 6:06:21 PM3/26/21
to Alex, TensorFlow Lite

Tiezhen Wang

unread,
Mar 29, 2021, 6:00:33 AM3/29/21
to Ivelin Ivanov, Yu-Cheng Ling, Alex, TensorFlow Lite
Hi Ivelin,

Model Maker makes it easier to transfer learning a model using your own dataset. The model can be deployed to a full range of devices including rpi. However the training process itself requires a workstation.

If you're looking for training on-device (without the need of a workstation), +Yu-Cheng Ling could be the best person to answer this question.

Thanks,
Tiezhen

--
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/CAP1uK1jbgOswCKAxuACZBv0s7-NvTFTXW8GU6m_hh-QAdpA-vg%40mail.gmail.com.

Ivelin Ivanov

unread,
Mar 29, 2021, 3:32:41 PM3/29/21
to TensorFlow Lite, Tiezhen Wang, Alex, TensorFlow Lite, Ivelin Ivanov, Yu-Cheng Ling
Yes, my question is certainly related to on-device training. If the training requires a workstation, that defeats the purpose of the concept and we can just use Keras/TF.

Yu-Cheng Ling

unread,
Apr 2, 2021, 6:04:57 PM4/2/21
to Ivelin Ivanov, TensorFlow Lite, Tiezhen Wang, Alex
Hi Ivelin,

Sorry for the delayed response. 

We (TFLite team) are actively working on supporting on-device training. 
We're currently implementing required infrastructure for on-device training in TFLite (e.g. supporting variables, gradient computation etc). 
Stay tuned and we will keep you posted in this group when we have something ready to use. 

Best,
YC

Ivelin Ivanov

unread,
Apr 2, 2021, 6:34:51 PM4/2/21
to Yu-Cheng Ling, TensorFlow Lite, Tiezhen Wang, Alex
Great, thanks for the update, YC. I will keep an eye for updates.

You are probably aware that the public TFLite roadmap was last updated a year ago (April 2020).
Would be great if you can share some of the key milestones on github. Maybe here?

That would provide better transparency and foster participation from the dev community. TFLite is still an open source project, correct?

Regards,

Ivelin

Shuangfeng Li

unread,
Apr 7, 2021, 1:27:04 AM4/7/21
to TensorFlow Lite, ivelin...@ambianic.ai, TensorFlow Lite, Tiezhen Wang, Alex, Yu-Cheng Ling
Ivelin, thanks for the suggestions, and TFLite team will update the roadmap doc shortly, and keep it updated more frequently.
Yeah, keeping all docs up-to-date regularly is not easy and we are working hard for that :-)
Reply all
Reply to author
Forward
0 new messages