Convert keras-yolo3 to onnx

TL;DR

-Original data learning procedure using Keras version implementation of YOLO v3 (as of June 24, 2020) HDF5 format (* .h5) created How to convert to ONNX format

Reference site

axinc-ai/yolov3-face onnx/keras-onnx

important point

Sometimes the converted onnx file could not be inferred. It can be inferred from the onnx file that was originally prepared.

The detailed conditions have not yet been identified.

1. Creating an environment

1.1. Source preparation

Clone with git from axinc-ai / yolov3-face

git clone https://github.com/axinc-ai/yolov3-face.git

Next, clone keras-yolo3 linked in yolov3-face.

cd .\yolov3-face
git clone https://github.com/qqwweee/keras-yolo3.git

1.2. Creating a python environment for conversion

Create a virtual environment "yolov3-face" with conda

conda create -n yolov3-face python=3.6 -y

Install the required modules. Since the detailed version is specified in axinc-ai / yolov3-face and it is not prepared in the conda repository, install it with pip.

conda activate yolov3-face
pip install tensorflow==1.13.2
pip install keras==2.2.4
pip install keras2onnx==1.5.1
pip install opencv-python
pip install pillow
pip install matplotlib

1.3. Creating a python environment for inference

Create virtual environment "yolov 3-face-inference" with conda

conda create -n yolov3-face-inference python=3.6 -y

Install the required modules. Since the detailed version is specified in axinc-ai / yolov3-face and it is not prepared in the conda repository, install it with pip.

conda activate yolov3-face-inference
pip install tensorflow==1.13.2
pip install onnxruntime
pip install keras==2.2.4
pip install pillow

2. Conversion and judgment (verified with the provided example)

First, verify with the prepared sample

2.1. Conversion

** Format **

python keras-yolo3-to-onnx.py <h5-model-filepath> <classes-filepath> <anchors-filepath> <onnx-filepath>

** Run **

cd .\keras-onnx

conda activate yolov3-face
python keras-yolo3-to-onnx.py ../model_data/logs/trained_weights_final.h5 ../model_data/face_classes.txt ../model_data/tiny_yolo_anchors.txt ../model_data/ax_face.onnx

2.2. Reasoning

** Format **

python inference.py <onnx-filepath> <classes-filepath> <targetimage-filepath> <outputimage-filepath>

** Run **

cd .\keras-onnx

conda activate yolov3-face-inference
python inference.py ../model_data/ax_face.onnx ../model_data/face_classes.txt ../images/couple.jpg output.jpg

3. For files processed by Original data learning procedure using Keras version implementation of YOLO v3 (as of June 24, 2020)

• Copy the file to the location described.

3.1. Conversion

cd .\keras-onnx

conda activate yolov3-face
python keras-yolo3-to-onnx.py ../model_data/yolo_logs/models/Step2_yolo_weight_mAP_best.h5 ../model_data/voc_classes.txt ../model_data/yolo_anchors.txt ../model_data/ax_yolov3.onnx

3.2. Reasoning

python inference.py ../model_data/ax_yolov3.onnx ../model_data/voc_classes.txt ../images/couple.jpg output.jpg

reference

keras-yolo3 + JetsonNano

https://qiita.com/rhene/items/b2a8ebe1f003e1107f63 http://mirai-tec.hatenablog.com/entry/2019/08/24/102129 http://mirai-tec.hatenablog.com/entry/2019/09/03/235156

YOLOv3(Darknet) + JetsonNano

https://soralab.space-ichikawa.com/2019/06/jetson-tx2-tensorrt-yolov3/ https://www.nakasha.co.jp/future/ai/vol2_yolov3nvidia_jetson_nano.html

TensorRT + Jetson Nano https://qiita.com/tsutof/items/f81d3900fa77d954ef39

Convert from kerass to ONNX

https://medium.com/axinc/yolov3-66c9b998c096

Other

https://qiita.com/agumon/items/114da6921c5dc4f7d7f9 https://github.com/zzh8829/yolov3-tf2 https://qiita.com/plseal/items/f493c67b2e810f2f876e http://mirai-tec.hatenablog.com/entry/2020/01/12/115546 https://rightcode.co.jp/blog/information-technology/tensorflow2-yolov3-run