Skip to content

Latest commit

 

History

History
 
 

ptq

README.md

This document describes the step-by-step to reproduce Yolo-v3 tuning result with Neural Compressor.

Prerequisite

1. Installation

Recommend python 3.6 or higher version.

# Install Intel® Neural Compressor
pip install neural-compressor

2. Install Intel Tensorflow

Check your python version and pip install 1.15.0 up3 from links as below:   
pip install https://storage.googleapis.com/intel-optimized-tensorflow/intel_tensorflow-1.15.0up3-cp36-cp36m-manylinux_2_12_x86_64.manylinux2010_x86_64.whl   
pip install https://storage.googleapis.com/intel-optimized-tensorflow/intel_tensorflow-1.15.0up3-cp37-cp37m-manylinux_2_12_x86_64.manylinux2010_x86_64.whl

Note: Supported Tensorflow versions please refer to Neural Compressor readme file.

3. Installation Dependency packages

cd examples/tensorflow/object_detection/yolo_v3/quantization/ptq
pip install -r requirements.txt

4. Downloaded Yolo-v3 model

git clone https://github.com/mystic123/tensorflow-yolo-v3.git
cd tensorflow-yolo-v3

5. Download COCO Class Names File

wget https://raw.githubusercontent.com/pjreddie/darknet/master/data/coco.names

6. Download Model Weights (Full):

wget https://pjreddie.com/media/files/yolov3.weights

7. Generate PB:

python convert_weights_pb.py --class_names coco.names --weights_file yolov3.weights --data_format NHWC --size 416 --output_graph yolov3.pb

8. Prepare Dataset

Automatic dataset download

Note: prepare_dataset.sh script works with TF version 1.x.

Run the prepare_dataset.sh script located in examples/tensorflow/object_detection/yolo_v3/quantization/ptq.

Usage:

cd examples/tensorflow/object_detection/yolo_v3/quantization/ptq
. prepare_dataset.sh

This script will download the train, validation and test COCO datasets. Furthermore it will convert them to tensorflow records using the https://github.com/tensorflow/models.git dedicated script.

Manual dataset download

Download CoCo Dataset from Official Website.

Get Quantized Yolo-v3 model with Neural Compressor

1.Config the yolo_v3.yaml with the valid cocoraw data path.

2.Run below command one by one.

Usage

cd examples/tensorflow/object_detection/yolo_v3/quantization/ptq
python infer_detections.py --input_graph /path/to/yolov3_fp32.pb --config ./yolo_v3.yaml --output_graph /path/to/save/yolov3_tuned3.pb

Finally, the program will generate the quantized Yolo-v3 model with relative 1% loss.