Intel® Neural Compressor Sample for TensorFlow*

Background

Low-precision inference can speed up inference obviously, by converting the fp32 model to int8 or bf16 model. Intel provides Intel® Deep Learning Boost technology in the Second Generation Intel® Xeon® Scalable Processors and newer Xeon®, which supports to speed up int8 and bf16 model by hardware.

Intel® Neural Compressor helps the user to simplify the processing to convert the fp32 model to int8/bf16.

At the same time, Intel® Neural Compressor will tune the quanization method to reduce the accuracy loss, which is a big blocker for low-precision inference.

Intel® Neural Compressor is released in Intel® AI Analytics Toolkit and works with Intel® Optimization of TensorFlow*.

Please refer to the official website for detailed info and news: https://github.com/intel/neural-compressor

Introduction

This is a demo to show an End-To-End pipeline to build up a CNN model to recognize handwriting number and speed up AI model by Intel® Neural Compressor.

Train a CNN AlexNet model by Keras and Intel Optimization for Tensorflow based on dataset MNIST.
Quantize the frozen PB model file by Intel® Neural Compressor to INT8 model.
Compare the performance of FP32 and INT8 model by same script.

We will learn the acceleration of AI inference by Intel AI technology:

Intel® Deep Learning Boost
Intel® Neural Compressor
Intel® Optimization for Tensorflow*

Code

Function	Code	Input	Output
Train a CNN AlexNet model	keras_tf_train_mnist.py	dataset: MNIST	fp32_frozen.pb
Quantize the frozen PB model file	inc_quantize_model.py	dataset: MNIST model: fp32_frozen.pb yaml: alexnet.yaml	alexnet_int8_model.pb
Test performance	profiling_inc.py	fp32_frozen.pb alexnet_int8_model.pb	32.json 8.json
Compare the performance	compare_perf.py	32.json 8.json	stdout/stderr log file fp32_int8_aboslute.png fp32_int8_times.png

run_inc_ft_mnist_sample.sh will call above python scripts to finish the demo.

Getting Started with Intel® DevCloud

This article assumes you are familiar with Intel® DevCloud environment. To learn more about working with Intel® DevCloud, please refer to Intel® DevCloud. Specifically, this article assumes:

You have an Intel® DevCloud account.
You are familiar with usage of Intel® DevCloud, like login by SSH client..
Developers are familiar with Python, AI model training and inference based on Tensorflow*.

Hardware Environment

This demo could be executed on any Intel CPU. But it's recommended to use 2nd Generation Intel® Xeon® Scalable Processors or newer, which include:

AVX512 intruction to speed up training AI model.
Intel® Deep Learning boost: Vector Neural Network Instruction (VNNI) to accelerate AI/DL Inference.

Setup Running Environment

Intel® DevCloud

We prepared the running enviroment by Conda in Intel® DevCloud for oneAPI. Activate it by:

conda activate /data/oneapi_workshop/INC

If you want to setup your own Conda environment, please refer to next chapter: Customer Server.

Customer Server

Set up own Conda enviroment in local server, cloud (including Intel® DevCloud):

Base on Intel® oneAPI AI Analytics Toolkit

Install Intel® oneAPI AI Analytics Toolkit.

For installation instructions, refer to Intel® oneAPI Toolkits Installation Guides.

Set up your Intel® oneAPI AI Analytics Toolkit environment.

Change the oneAPI installed path in the following command, according to your installation.

In this case, we use /opt/intel/oneapi.

source /opt/intel/oneapi/setvars.sh

Create Conda Envrionment: user_tensorflow

./set_env.sh

Create a new kernel for the Jupyter notebook based on your activated conda environment (Optional).

conda install ipykernel
python -m ipykernel install --user --name usr_tensorflow

Install from Scrach by Conda

Create Conda Envrionment: user_tensorflow in Intel channel.

conda create -n user_tensorflow -c intel python=3.9 -y
conda activate user_tensorflow
conda install -n user_tensorflow -c intel tensorflow python-flatbuffers -y
conda install -n user_tensorflow -c intel neural-compressor -y
conda install -n user_tensorflow runipy notebook -y

Create a new kernel for the Jupyter notebook based on your activated conda environment (Optional).

conda install ipykernel
python -m ipykernel install --user --name usr_tensorflow

Update Script

Edit run_inc_ft_mnist_sample.sh to set the Conda enviroment according to above result: /data/oneapi_workshop/INC or user_tensorflow.

vi run_inc_ft_mnist_sample.sh

...
conda activate user_tensorflow
or
conda activate /data/oneapi_workshop/INC
...

Run in Intel® DevCloud for oneAPI

Run in Jupyter Notebook in Intel® DevCloud for oneAPI

Please open inc_sample_for_tensorflow.ipynb in Jupyter Notebook.

Following the guide to run this demo.

Run in SSH Login Intel® DevCloud for oneAPI

This demo will show the obviously acceleration by VNNI. In Intel® DevCloud, please choose compute node with the property 'clx' or 'icx' or 'spr' which support VNNI.

Job Submit

!qsub run_inc_ft_mnist_sample.sh -d `pwd` -l nodes=1:icx:ppn=2
28029.v-qsvr-nda.aidevcloud

Note, please run above command in login node. There will be error as below if run it on compute node:

qsub: submit error (Bad UID for job execution MSG=ruserok failed validating uXXXXX/uXXXXX from s001-n054.aidevcloud)

Check job status

qstat

After the job is over (successfully or fault), there will be log files:

run_inc_ft_mnist_sample.sh.o28029
run_inc_ft_mnist_sample.sh.e28029

Check Result

Check Result in Log File

Check the result in log file: run_inc_ft_mnist_sample.sh.o28029:

!tail -23 run_inc_ft_mnist_sample.sh.o1842253


Model            FP32                     INT8                    
throughput(fps)  572.4982883964987        3030.70552731285        
latency(ms)      2.8339174329018104       2.128233714979522       
accuracy(%)      0.9799                   0.9796                  

Save to fp32_int8_aboslute.png

Model            FP32                     INT8                    
throughput_times 1                        5.293824608282245       
latency_times    1                        0.7509864932092611      
accuracy_times   1                        0.9996938463108482      

Save to fp32_int8_times.png
Please check the PNG files to see the performance!
This demo is finished successfully!
Thank you!

########################################################################
# End of output for job 1842253.v-qsvr-1.aidevcloud
# Date: Thu 27 Jan 2022 07:05:52 PM PST
########################################################################

...

We will see the performance and accuracy of FP32 and INT8 model. The performance could be obviously increased if running on Xeon with VNNI.

Check Result in PNG file

The demo creates figure files: fp32_int8_aboslute.png, fp32_int8_times.png to show performance bar. They could be used in report.

Copy files from DevCloud in host:

scp devcloud:~/intel_neural_compressor_sample/*.png ./

Run in Customer Server or Cloud

Note, it's recommended to use 2nd Generation Intel® Xeon® Scalable Processors or newer to get better performance improvement.

Run in Jupyter Notebook

Please open inc_sample_for_tensorflow.ipynb in Jupyter Notebook.

Following the guide to run this demo.

Run in SSH Login

run_inc_ft_mnist_sample.sh

Check Result

Check Result in Output

...

Model            FP32                     INT8                    
throughput(fps)  572.4982883964987        3030.70552731285        
latency(ms)      2.8339174329018104       2.128233714979522       
accuracy(%)      0.9799                   0.9796                  

Save to fp32_int8_aboslute.png

Model            FP32                     INT8                    
throughput_times 1                        5.293824608282245       
latency_times    1                        0.7509864932092611      
accuracy_times   1                        0.9996938463108482      

Save to fp32_int8_times.png
Please check the PNG files to see the performance!
This demo is finished successfully!
Thank you!

########################################################################
# End of output for job 1842253.v-qsvr-1.aidevcloud
# Date: Thu 27 Jan 2022 07:05:52 PM PST
########################################################################

...

Check Result in PNG file

The demo creates figure files: fp32_int8_aboslute.png, fp32_int8_times.png to show performance bar. They could be used in report.

License

Code samples are licensed under the MIT license. See License.txt for details.

Third party program Licenses can be found here: third-party-programs.txt

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

README.md

README.md

Intel® Neural Compressor Sample for TensorFlow*

Background

Introduction

Code

Getting Started with Intel® DevCloud

Hardware Environment

Setup Running Environment

Intel® DevCloud

Customer Server

Base on Intel® oneAPI AI Analytics Toolkit

Install from Scrach by Conda

Update Script

Run in Intel® DevCloud for oneAPI

Run in Jupyter Notebook in Intel® DevCloud for oneAPI

Run in SSH Login Intel® DevCloud for oneAPI

Job Submit

Check job status

Check Result

Check Result in Log File

Check Result in PNG file

Run in Customer Server or Cloud

Run in Jupyter Notebook

Run in SSH Login

Check Result

Check Result in Output

Check Result in PNG file

License

Files

README.md

Latest commit

History

README.md

File metadata and controls

Intel® Neural Compressor Sample for TensorFlow*

Background

Introduction

Code

Getting Started with Intel® DevCloud

Hardware Environment

Setup Running Environment

Intel® DevCloud

Customer Server

Base on Intel® oneAPI AI Analytics Toolkit

Install from Scrach by Conda

Update Script

Run in Intel® DevCloud for oneAPI

Run in Jupyter Notebook in Intel® DevCloud for oneAPI

Run in SSH Login Intel® DevCloud for oneAPI

Job Submit

Check job status

Check Result

Check Result in Log File

Check Result in PNG file

Run in Customer Server or Cloud

Run in Jupyter Notebook

Run in SSH Login

Check Result

Check Result in Output

Check Result in PNG file

License