TF2.0 the fate of `contrib.quantize` project

[Sergey Mironov]

Hi, all! I'd like to ask about plans for contrib.quantize project in TensorFlow 2.0.

From StackOverflow I found the follwowing link

https://github.com/tensorflow/community/blob/master/rfcs/20180907-contrib-sunset.md#list-of-projects

saying that the owner of `contrib.quantize` is Suharshs, but no fate is specified.

Currently I got an unrelated error on attempt to import anything from `tensorflow.contrib`. Could anybody provide me with some information?

Thanks in advance,

Sergey

PS

I'm using TF 2.0.0-rc1.

[Paige Bailey]

Adding +Suharsh Sivakumar and +Raziel Alvarez, RE: quantization.

--
You received this message because you are subscribed to the Google Groups "TensorFlow Developers" group.
To unsubscribe from this group and stop receiving emails from it, send an email to developers+...@tensorflow.org.
To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/developers/218390d0-50c4-4e31-9808-39540f965675%40tensorflow.org.

--

• • • •

Paige Bailey    Product Manager (TensorFlow) @DynamicWebPaige webp...@google.com

[Sergey Mironov]

Thanks for adding people.

Update: I found some quantization code in the `core` codebase and tried to re-produce some tests, but

1. I see the following deprecation warning

```

WARNING:tensorflow:From /workspace/mironov/hbtest/fq_transformer_bug.py:40: do_quantize_training_on_graphdef (from tensorflow.python.pywrap_tensorflow_internal) is deprecated and will be removed in a future version.                        

Instructions for updating:

GraphDef quantized training rewriter is deprecated in the long term

```

2. Test `testQuantizeTraining` doesn't seem to work for me (the last assert is triggered, I conculde that `do_quantize_training_on_graphdef` didn't modify the graph in any way)

I'm interested in fake-quantization, either automatic (via graph rewrite) or manual (as separate operations). Could you please share the plans and views you have regarding this subject? 

Thanks in advance,

Sergey Mironov

понедельник, 23 сентября 2019 г., 17:55:41 UTC+3 пользователь Paige Bailey написал:

Adding +Suharsh Sivakumar and +Raziel Alvarez, RE: quantization.

On Mon, Sep 23, 2019 at 7:17 AM Sergey Mironov <grr...@gmail.com> wrote:

Hi, all! I'd like to ask about plans for contrib.quantize project in TensorFlow 2.0.

From StackOverflow I found the follwowing link

https://github.com/tensorflow/community/blob/master/rfcs/20180907-contrib-sunset.md#list-of-projects

saying that the owner of `contrib.quantize` is Suharshs, but no fate is specified.

Currently I got an unrelated error on attempt to import anything from `tensorflow.contrib`. Could anybody provide me with some information?

Thanks in advance,

Sergey

PS

I'm using TF 2.0.0-rc1.

--
You received this message because you are subscribed to the Google Groups "TensorFlow Developers" group.

To unsubscribe from this group and stop receiving emails from it, send an email to devel...@tensorflow.org.

To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/developers/218390d0-50c4-4e31-9808-39540f965675%40tensorflow.org.

[Piyush Datta]

Hi Sergey,

I’m not sure about the fate of contrib.quantize, however the contrib folder was removed on September 18. 

That’s why you aren’t able to import anything. This is for master and 2.X branches. 1.X branches still have contrib folder.

I’ve forwarded you the email regarding the removal of the folder.

Best Regards,

Piyush Datta

To unsubscribe from this group and stop receiving emails from it, send an email to developers+...@tensorflow.org.
To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/developers/6234c60e-3994-4894-954d-1e397ff6ebd4%40tensorflow.org.

--

Best Regards,

Piyush Datta

[王振华]

Hi guys,

I have raised same concern before, and Suharshs told that there will be a Keras API in 2.x to support quantization aware training.

Hoping this helpful :)

Cheers

Zhenhua

在 2019年9月24日星期二 UTC+8上午12:07:59，Piyush Datta写道：

To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/developers/6234c60e-3994-4894-954d-1e397ff6ebd4%40tensorflow.org.

[Sergey Mironov]

Zhenhua, Piyush, thanks for the information.

I'm looking forward to trying new Keras implementation. Of all kinds of quantization, I am more interested in quantization-aware training since we look at less-than-8 bits designs. If you need some help or interested in early testing of Keras implementation - please let me know. Post-training quantization is probably out of my scope right now.

Model Optimization Toolkit is fine, I see quntization things in it's roadmap and I'd like to know how does it relate to the Keras approach. If I understand the intentions correctly, model-optimization toolkit is designed to be a kind of automatic tool while Keras usually provides more fine-grained functional-style API. Please correct me if there is a better explanation.

Regards,

Sergey Mironov

вторник, 24 сентября 2019 г., 9:08:43 UTC+3 пользователь 王振华 написал:

[Suharsh Sivakumar]

Hi Sergey,

The Keras based API will be released as part of the model optimization toolkit. Will keep you posted!

-Suharsh

[Sergey Mironov]

> It's the same.

Silly me :) OK, thanks again!

вт, 24 сент. 2019 г. в 21:13, Raziel Alvarez <raz...@google.com>:

"Model Optimization Toolkit is fine, I see quntization things in it's roadmap and I'd like to know how does it relate to the Keras approach."

It's the same.

As soon as it's ready there will be a blog post in the TensorFlow blog, and twitter and other TF user groups.

Thanks