Quantize Inception V3 by Intel® Extension for Tensorflow* on Intel® Xeon®

Background

Intel® Extension for TensorFlow* provides quantization feature by cooperating with Intel® Neural Compressor and oneDNN Graph. It will provide better quantization: better performance and accuracy loss under control.

Intel® Neural Compressor executes the calibration process to output the QDQ quantization model, which inserts Quantize and Dequantize layers to includes help information for quantization.

When you use Intel® Extension for Tensorflow* to execute the inference of this model, oneDNN Graph will be called to quantize and optimize the model. Then the quantized model will be executed by Intel® Extension for Tensorflow* and accelerated by Intel® Deep Learning Boost or Intel® Advanced Matrix Extensions on Intel® Xeon® processors.

Introduction

The example shows an end-to-end pipeline:

  1. Train an Inception V3 model with a flower photo dataset by transfer learning.

  2. Execute the calibration by Intel® Neural Compressor.

  3. Quantize and accelerate the inference by Intel® Extension for Tensorflow* for GPU and CPU.

Configuration

Intel® Extension for Tensorflow* Version

Install Intel® Extension for Tensorflow* >= 2.13.0.0 for this feature.

Enable oneDNN Graph

By default, oneDNN Graph is enabled in Intel® Extension for Tensorflow* for INT8 models.

Enable it explicitly by:

  import os
  os.environ["ITEX_ONEDNN_GRAPH"] = "1"

Disable Constant Folding Function

We need to disable Constant Folding function in 2 stages:

  1. Intel® Neural Compressor creates QDQ quantization model.

  2. Intel® Extension for Tensorflow* executes the oneDNN Graph quantization path.

There are 2 methods to configure:

a. Environment Variable

export ITEX_TF_CONSTANT_FOLDING=0

b. Python API

from tensorflow.core.protobuf import rewriter_config_pb2

infer_config = tf.compat.v1.ConfigProto()
infer_config.graph_options.rewrite_options.constant_folding = rewriter_config_pb2.RewriterConfig.OFF

session = tf.compat.v1.Session(config=infer_config)
tf.compat.v1.keras.backend.set_session(session)

Hardware Environment

Support: Intel® Xeon® CPU & Intel® Data Center Flex Series GPU.

CPU

It’s recommended to run the example on the Intel® Xeon® processors, which supports Intel® Deep Learning Boost or Intel® Advanced Matrix Extensions.

Without the hardware features above for AI workloads, the performance speedup with FP32 will not be increased much.

Check Intel® Deep Learning Boost

In Linux, run command:

lscpu | grep vnni

You are expected to see avx_vnni and avx512-vnni, otherwise your processors do not support Intel® Deep Learning Boost.

Check Intel® Advanced Matrix Extensions

In Linux, run command:

lscpu | grep amx

You are expected to see amx_bf16 and amx_int8, otherwise your processors do not support Intel® Advanced Matrix Extensions.

GPU

Support: Intel® Data Center Flex Series GPU.

Local Server

Install the GPU driver and oneAPI packages by referring to Intel GPU Software Installation.

Intel® DevCloud

If you have no CPU support Intel® Deep Learning Boost or Intel® Advanced Matrix Extensions or no Intel GPU support INT8, you could register on Intel® DevCloud and try this example on an second generation Intel® Xeon based processors or newer. To learn more about working with Intel® DevCloud, refer to Intel® DevCloud

Running Environment

  1. Install Python versions >=3.8 and versions <=3.10 supported by Intel® Extension for Tensorflow*.

  2. Create the running Python Virtual environment env_itex.

bash pip_set_env.sh

  1. Activate

source env_itex/bin/activate

Startup Jupyter Notebook

  1. Startup

bash run_jupyter.sh

...
http://xxx.yyy.com:8888/xxxxxxxx

  1. Open the link outputted by Jupyter Notebook in your browser.

  2. Choose and open the quantize_inception_v3.ipynb in Jupyter Notebook.

Set the kernel to “env_itex”.

Execute the code as the guide.

License

Code samples are licensed under the MIT license.