Federated Learning

[Agnik Banerjee]

Does anybody know of any frameworks that implement Federated Learning using TFLite on Android or iOS?

[Haoliang Zhang]

+Chunxiang (Jake) Zheng Could you help advise?

On Tue, Jun 28, 2022 at 12:06 PM Agnik Banerjee <agnik....@gmail.com> wrote:

Does anybody know of any frameworks that implement Federated Learning using TFLite on Android or iOS?

--
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/96f715cd-c3c9-43af-9953-67cf64607ef6n%40tensorflow.org.

--

Best,

Haoliang

[Lingchuan Meng]

There is a paper called "On-Device Federated Learning With Flower" that uses the TFLite model personalization support in their implementation. But now that on-device training in TFLite has been introduced, FL frameworks should probably use the new flow. I recall seeing "support for FL" on TFLite roadmap at some point. Maybe the team members can provide a concrete answer.

[Haoliang Zhang]

Hi Lingchuan,

FL will be supported in the future. TFF is one way to write Federated programs in Tensorflow, and the client-side execution runtime will be TF Lite. However, there are still some core libraries associated with orchestrating the training between devices/server are still not open-sourced. We will need to wait for that to be open-sourced in the future, thanks for your interest!

--

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/8de5c5d8-f9a2-44f8-b454-31730d3a9973n%40tensorflow.org.

--

Best,

Haoliang

[Lingchuan (LC) Meng]

Hi Haoliang,

Thank you for the answer. Good to know that the technology and implementations are both there, mostly pending on the open-source decision before they become available publicly.

What kind of models will get the FL support in TFLite, e.g., FP32 vs. fully quantized? Thanks.

[Pannag Sanketi]

Have you looked at Bella? 

To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/tflite/CAO5Df84kifZL_0qDwe5siEqxCGAtHfSz-RFmU925w%2BSRQ_S9pA%40mail.gmail.com.

[Haoliang Zhang]

I think only FP32 will be supported first. 

--

Best,

Haoliang

[Pannag Sanketi]

I meant Brella. I was part of the project earlier but not anymore. I believe go/brella should still be valid.

[Haoliang Zhang]

Hi Pannag,

The user is coming from outside of Google. So they can not use Brella directly (or anything non open-sourced).

--

Best,

Haoliang

[Pannag Sanketi]

Oops, sorry, I did not realize this was an external request. Apologies! 

[Haoliang Zhang]

No worries! For open-source plan, maybe @Chunxiang (Jake) Zheng can give more accurate estimates.

--

Best,

Haoliang

[kwang...@gmail.com]

May I ask when "FL support on TFLite" is available and what features are supported?

We've been waiting for it for years.

 

Thanks in advance,

Kwangkee

To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/tflite/CAM37U-PGZeT3%2B%2B25c%2B535DuV4%2Bx0%2BSaE3T8%2BfaEV6%2BhsXFy%3D9g%40mail.gmail.com.

[Haoliang Zhang]

Hi, 

Can I know if this will be used for your company's product? Or is it mainly used for personal development?

--

Best,

Haoliang

[kwang...@gmail.com]

Thanks for your interest.

Let me share the brief of our project in progress:

 

ㅇ Development of adaptive federated learning technology that is robust against statistical heterogeneity of data and systemic heterogeneity of devices, and can optimally reflect the characteristics and application requirements of individual devices and users

- Adaptive federated learning technology that can support personalized learning models along with global learning models

- Extended federated learning technology that adaptively responds to continuous environmental changes and efficiently supports knowledge transfer between heterogeneous models

- Applying adaptive federated learning technology to real scenarios of Unsupervised Learning, Supervised Learning, and Reinforcement Learning to present/verify the effectiveness and excellence of the technology, and secure performance and efficiency that can be utilized in real-world applications

 

In short, immediate "FL support on TFLite" is indispensable for our project to move forward.

[Haoliang Zhang]

+Mark Sherwood +Yang Lu+Yu-Cheng Ling +Ram Iyengar 

Thank you for your detailed introduction of your project. Supporting FL for Tensorflow Lite users will be our top priority, and we are still working on finalizing the concrete roadmap. We will keep you updated once we have more to share.

Meanwhile, I think there is still some gap to map those high level objectives to product requirements, I'm not sure what's the best way for future collaboration, I will defer it to our PM Mark/Yang to see how to better move forward with this.

Thanks!

--

Best,

Haoliang

[Chunxiang (Jake) Zheng]

+Krzysztof Ostrowski 

[kwang...@gmail.com]

It's a pity that, as you may agree, on-device training support for TFLite is critical for such scenarios as Federated Learning, Meta Learning, and Continual learning to be introduced into the industry.

 

Sure, feel free to contact me.

Kwangkee

 

From: Mark Sherwood <marksh...@google.com>
Sent: Saturday, July 16, 2022 10:45 AM
To: Chunxiang (Jake) Zheng <chunxia...@google.com>
Cc: Haoliang Zhang <haol...@google.com>; Krzysztof Ostrowski <ostr...@google.com>; kwang...@gmail.com; Yang Lu <yangj...@google.com>; Yu-Cheng Ling <ycl...@google.com>; Ram Iyengar <ramiy...@google.com>; Lingchuan (LC) Meng <lingch...@gmail.com>; agnik....@gmail.com <agnik....@gmail.com>; Pannag Sanketi <psan...@google.com>; TensorFlow Lite <tfl...@tensorflow.org>
Subject: Re: Federated Learning

 

Hi Kwangkee,

Thanks for reaching out and expressing interest in TFLite and On Device Training / Federated Learning. Unfortunately at this time we cannot make any commitments for timelines or even availability of future open source releases or products. If we get to a point where we are looking for community input or interviews to help guide product roadmaps, would you be interested in participating? If so, please let me know your company information and contact information and I can note that for potential future opportunities.

Thanks,
Mark