Quantized 3D Convolution support in TensorFlow Lite

38 views
Skip to first unread message

Robert Kalmar

unread,
Jan 18, 2023, 4:27:06 AM1/18/23
to TensorFlow Lite
Hello,
recently the Conv3D operator was added to TensorFlow Lite Operator Set. However so far only FP32 variant is available. No Quantization support for this operator. Also I did not noticed any activity regarding to this on TensorFlow's master branch nor RFC.

I would like to ask you regarding the plans for Quantization support for Conv3D. Is on the roadmap?

Thank you,
Robert Kalmar

Haoliang Zhang

unread,
Jan 21, 2023, 3:35:30 PM1/21/23
to Robert Kalmar, TensorFlow Lite
Hi Robert.

Currently there is no plan adding quantization support for Conv3D. 

--
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/f9d7f0ec-ed06-483e-8ec9-28e0cfcfc9a3n%40tensorflow.org.


--
Best,
Haoliang

Elena Zhelezina

unread,
Jan 24, 2023, 7:48:45 AM1/24/23
to TensorFlow Lite, haol...@google.com, TensorFlow Lite, robert...@nxp.com
Hello,

FYI, there is this PR https://github.com/tensorflow/tensorflow/pull/51527 that adds Conv3D support.

Regards,
Elena.

Robert Kalmar

unread,
Jan 25, 2023, 10:08:53 AM1/25/23
to TensorFlow Lite, elena.z...@gmail.com, haol...@google.com, TensorFlow Lite, Robert Kalmar
Thank you,
Elena and Haoliang for the response.

Haoliang, how can I understand there is no plan adding the quantization support for Conv3D. You mean the TensorFlow Core Development team/Google is not planning it, but contribution is welcome. Or no plan at all, which impacts also the eventual third party contribution in this area?

Elena, can you please briefly describe the status of the PR. I see it is open since August 2021.

Best Regards,
Robert

Haoliang Zhang

unread,
Jan 25, 2023, 5:52:53 PM1/25/23
to Robert Kalmar, Jaesung Chung, Songyi Han, TensorFlow Lite, elena.z...@gmail.com
For the plan of quantization support, I will let @Jaesung Chung  or @Songyi Han to give more context on it.
--
Best,
Haoliang

YoungSeok Yoon

unread,
Feb 8, 2023, 10:18:21 AM2/8/23
to Haoliang Zhang, Robert Kalmar, Jaesung Chung, Songyi Han, TensorFlow Lite, elena.z...@gmail.com
Hi Robert, Elena,

Thanks for your interests, and apologies for the late reply.
We will resume the PR review process soon and get back to you.

In the meantime, have you tried patching in the existing PR for your use case?
Would that unblock you for the moment?

Thanks,
YoungSeok

Robert Kalmar

unread,
Feb 17, 2023, 3:07:48 AM2/17/23
to TensorFlow Lite, youngs...@google.com, Robert Kalmar, jae...@google.com, Songyi Han, TensorFlow Lite, elena.z...@gmail.com, haol...@google.com
Hello YoungSeok,
thank you for the reply. I tried out the solution from the patch and work for us for now. It unblocks us for the moment, but also we would appreciate if the review process can continue and eventually the support for Quantized 3D Convolution will be part of the future TensorFlow releases.

By the way, Haoliang, Jaesung, Songyi, I tried you reach you out outside of community mailing list. Did you get my message? Are you reachable on these email addresses also?

Best Regards,
Robert
Reply all
Reply to author
Forward
0 new messages