Neural Coder enable support of Keras script-based model

[kaikaiyao]

Type of Change

Feature addition: support Keras script-based model as keras_inc

Description

• Enabled support of auto-quantizing Keras script-based model
• Also enabled inc_auto as a default optimization for launcher and for API: our domain detector will detect whether the code is PT/TF/Keras/ONNX and choose default feature accordingly:
  • PyTorch: default dynamic quantization scheme
  • TensorFlow/Keras model based: default TensorFlow quantization scheme
  • Keras script-based: default Keras script-based static quantization scheme
  • ONNX: default dynamic quantization scheme
• Also enabled "enable_only" -e mode/argument in neural_coder launcher for testing
• Update docs accordingly

Expected Behavior & Potential Risk

Able to enable INC on Keras script-based model, for example this code

The current enabling pattern design is in accordance with the Keras example which was merged a few days ago, with key detection of model = keras.Sequential(xxx):

from neural_compressor.quantization import fit
from neural_compressor.config import PostTrainingQuantConfig
from neural_compressor.experimental import common
config = PostTrainingQuantConfig(backend='itex')
quantized_model = fit(MODEL_NAME, conf=config, calib_dataloader=DATALOADER_NAME, eval_func=eval_func)

Note that user needs to define a dataloader and eval_func as in the example.

How has this PR been tested?

Locally, can test on any machine with

# installation
git clone https://github.com/intel/neural-compressor.git
cd neural-compressor
git checkout neural_coder_enable_keras
python setup.py develop
cd ..

# test
python -m neural_coder -e neural-compressor/neural_coder/examples/keras/mnist.py

Dependency Change?

None

[kaikaiyao]

@chensuyue hi Suyue, do you know why these checks are hang? Another PR is also same status. Thanks!

[chensuyue]

/Azurepipeline run

[azure-pipelines]

Azure Pipelines successfully started running 2 pipeline(s).