Add a feature of UniformQDQ - support CV/NLP model's OPs, includingConv, DepthwiseConv2D, MatMul, etc. Additional op support be added upon request.

[qgao007]

Type of Change

A feature of UniformQDQ

Description

A feature of UniformQDQ

• support CV/NLP model's OPs, includingConv, DepthwiseConv2D, MatMul, etc. Additional op support be added upon request.

Expected Behavior & Potential Risk

Generate a UniformQDQ int8 model. No foreseeable risk.

How has this PR been tested?

This PR is tested on selected models with representative ops from image-classfication, object detection, and nlp.

Dependency Change?

Need Tensorflow with qunit8-support.
https://github.com/intel-innersource/frameworks.ai.tensorflow.private-tensorflow/tree/mabuzain/add-quint8-support

[thuang6]

one general question, I saw the changes in both 2.x and 3.x folders, so this PR enables UniformQDQ supported in both 2.x and 3.x API?

[qgao007]

one general question, I saw the changes in both 2.x and 3.x folders, so this PR enables UniformQDQ supported in both 2.x and 3.x API?

Thanks @thuang6 for the question! We ran it in new pythonic 3.x API, assuming compatibility with 2.x. We will confirm this shortly with @nhatleSummer22

[qgao007]

loop in @ashahba and @mahmoud-abuzaina also.

[thuang6]

one general question, I saw the changes in both 2.x and 3.x folders, so this PR enables UniformQDQ supported in both 2.x and 3.x API?

Thanks @thuang6 for the question! We ran it in new pythonic 3.x API, assuming compatibility with 2.x. We will confirm this shortly with @nhatleSummer22

new feature is only expected to 3.x API.

Please note, 3.x has separated packaging for each framework, TF 3.x packing is listed below, which only contains common & tensorflow folder under neural_compressor. If the function relies on code in other 2.x folders, it does not work.

"neural_compressor_tf": {
"project_name": "neural_compressor_tf",
"include_packages": find_packages(
include=[
"neural_compressor.common",
"neural_compressor.common.",
"neural_compressor.tensorflow",
"neural_compressor.tensorflow.",
],
),
"package_data": {"": ["*.yaml"]},
"install_requires": fetch_requirements("requirements_tf.txt"),
},

[mahmoud-abuzaina]

Supporting only 3.x API is fine, we don't really need to depend on 2.x API.