qiyuangong / auto_optimization

OpenVINO auto_optimization

OpenVINO

This repo has been merged into Analytics-Zoo.

Analytics Zoo provides a unified analytics + AI platform that seamlessly unites Spark, TensorFlow, Keras, BigDL and OpenVINO programs into an integrated pipeline; the entire pipeline can then transparently scale out to a large Hadoop/Spark cluster for distributed training or inference.

Basic Requirements

1. Python & TensorFlow
2. OpenVINO 2018 installed

Basic workflow of OpenVINO

Basic workflow of OpenVINO optimization and calibration.

(TensorFlow) -optimization-> IR model -calibration-> int8 IR model

Note that OpenVINO int8 IR promises higher performance than IR model (FP32) by sacrificing a few precision (1% by default).

Both IR and int8 IR can be loaded by OpenVINO, and make prediction. User should ensure that optimization and calibration are correctly performed.

Auto Calibration

Assume you have OpenVINO IR model (converted from Caffe or TensorFlow), and you want to convert it into int8 model. To achive that, you can use OpenVINO calibration_tool or use our auto_calibration tool, which provides simpler API based on calibration_tool.

Basic Usage

python auto_calibration.py -m {OpenVINO IR model, *.xml} -i {validation data}

API usage

python auto_calibration.py -h

Validation Data The key of calibration is a carefully prepared validation data. For different models, you should prepare different validation data. Currently, our auto_calibration tool only support image classification and object detection models and validation data.

1. Image Classification Two kinds of validation dir are supported:

• Image in label dir

+-- 0
|   +-- Image_001.jpg
|   +-- Image_003.jpg
+-- 1
|   +-- Image_002.jpg
|   +-- Image_004.jpg

• val.txt with images

+-- val.txt
+-- Image_001.jpg
+-- Image_002.jpg
+-- Image_003.jpg
+-- Image_004.jpg

val.txt contains image path and label, separated by space.

Image_001.jpg 0
Image_002.jpg 1
Image_003.jpg 0
Image_004.jpg 1

2. Object Detection For OpenVINO 2018, only VOC format is supported for Object Detection. We simplify the dir structure:

+-- classes.txt
+-- Image_001.jpg
+-- Image_002.jpg
+-- Image_003.jpg
+-- Image_004.jpg
+-- Image_001.xml
+-- Image_002.xml
+-- Image_003.xml
+-- Image_004.xml

Herein *.xml is annotation files for images. classes.txt is the labels. Example of classes.txt is shown below:

none_of_the_above 0
aeroplane 1
bicycle 2
bird 3
boat 4
bottle 5
bus 6
car 7
cat 8
chair 9
cow 10
diningtable 11
dog 12
horse 13
motorbike 14
person 15
pottedplant 16
sheep 17
sofa 18
train 19
tvmonitor 20

References

1. OpenVINO
2. Analytics-Zoo
3. TensorFlow