Converting Yolo v3 models to TensorFlow and OpenVINO(IR) models - Sudhakar17/darknet GitHub Wiki

To use Yolo v3 model on OpenVINO framework, you should do 2 steps:

1. Convert yolov3.cfg/yolov3.weights to TensorFlow model frozen_darknet_yolov3_model.pb
2. Convert frozen_darknet_yolov3_model.pb to OpenVINO model frozen_darknet_yolov3_model.xml/.bin/.mapping

More: https://docs.openvinotoolkit.org/latest/_docs_MO_DG_prepare_model_convert_model_tf_specific_Convert_YOLO_From_Tensorflow.html

Converting a Yolo v3 model: Darknet -> TensorFlow

• Download this repository: https://github.com/mystic123/tensorflow-yolo-v3/archive/ed60b9087b04e1d9ca40f8a9d8455d5c30c7c0d3.zip

• un-pack it:

sudo apt-get install unzip
unzip file.zip -d tensorflow-yolo-v3
cd tensorflow-yolo-v3

• Download default Yolo v3 model: https://pjreddie.com/media/files/yolov3.weights

• Download coco.names file: https://raw.githubusercontent.com/AlexeyAB/darknet/master/data/coco.names

• run this command python3 convert_weights_pb.py --class_names coco.names --data_format NHWC --weights_file yolov3.weights

for tiny: python3 convert_weights_pb.py --class_names coco.names --data_format NHWC --weights_file yolov3-tiny.weights --tiny

• you will get TensorFlow model frozen_darknet_yolov3_model.pb

Converting a Yolo v3 model: TensorFlow -> OpenVINO(IR)

After converting the Darknet Yolo v3 model to the TensorFlow model, you can convert it to the Openvino model.

• Install OpenVINO: https://software.intel.com/en-us/openvino-toolkit/choose-download

• Put these files to the one directory:

  • frozen_darknet_yolov3_model.pb - TensorFlow model that you got in previous stage
  • <OPENVINO_INSTALL_DIR>/deployment_tools/model_optimizer/extensions/front/tf/yolo_v3.json
  • <OPENVINO_INSTALL_DIR>/deployment_tools/model_optimizer/mo_tf.python

• Run the command: python3 mo_tf.py -b 1 --input_model ./frozen_darknet_yolov3_model.pb --tensorflow_use_custom_operations_config ./yolo_v3.json --data_type FP16

• You will get OpenVINO model that can be run on CPU, GPU, VPU (MyriadX), FPGA:

  • frozen_darknet_yolov3_model.xml - model structure
  • frozen_darknet_yolov3_model.bin - model weights
  • frozen_darknet_yolov3_model.mapping - mapping file

Run IR-OpenVINO model on CPU or VPU (Myriad X)

1. Build C++ examples - run: <OPENVINO_INSTALL_DIR>/inference_engine/samples/build_samples.sh

   • cd /user/home/inference_engine_samples_build/intel64/Release

   • on CPU: ./object_detection_demo_yolov3_async -i ./test.mp4 -m ./frozen_darknet_yolov3_model.xml -d CPU

   • on VPU: ./object_detection_demo_yolov3_async -i ./test.mp4 -m ./frozen_darknet_yolov3_model.xml -d MYRIAD

2. By using Python code:

   • go to <OPENVINO_INSTALL_DIR>/inference_engine/samples/python_samples

   • python3 object_detection_demo_yolov3_async.py -i test.mp4 -m ./frozen_darknet_yolov3_model.xml -d CPU

   • read: https://docs.openvinotoolkit.org/latest/_inference_engine_ie_bridges_python_sample_object_detection_demo_yolov3_async_README.html

---

• More: https://docs.openvinotoolkit.org/latest/_inference_engine_samples_object_detection_demo_yolov3_async_README.html

• Also read: https://github.com/PINTO0309/OpenVINO-YoloV3

---

Run Yolo v3 models on OpenCV-dnn with OpenVINO DL-IE backend:

• Install OpenCV by using these bash-commands: https://github.com/opencv/opencv/wiki/Intel's-Deep-Learning-Inference-Engine-backend

cmake -D CMAKE_BUILD_TYPE=RELEASE \
 -D WITH_INF_ENGINE=ON \
 -D ENABLE_CXX11=ON \
 -D BUILD_EXAMPLES=OFF \
 -D WITH_FFMPEG=ON \
 -D WITH_V4L=OFF \
 -D WITH_LIBV4L=ON \
 -D OPENCV_ENABLE_PKG_CONFIG=ON \
 -D BUILD_TESTS=OFF \
 -D BUILD_PERF_TESTS=OFF \
 -D INF_ENGINE_LIB_DIRS="/opt/intel/openvino/deployment_tools/inference_engine/lib/intel64" \
 -D INF_ENGINE_INCLUDE_DIRS="/opt/intel/openvino/deployment_tools/inference_engine/include" \
 -D CMAKE_FIND_ROOT_PATH="/opt/intel/openvino/" \
 ..

make -j8
sudo make install

• Compile this sample by using make command: ocv_yolov3.zip

  Or use original sample: https://github.com/opencv/opencv/blob/master/samples/dnn/object_detection.cpp

• Download yolov3-tiny.cfg & yolov3-tiny.weights files: https://pjreddie.com/darknet/yolo/

  • https://raw.githubusercontent.com/pjreddie/darknet/master/cfg/yolov3-tiny.cfg
  • https://pjreddie.com/media/files/yolov3-tiny.weights

• Run command: ./ocv_yolo --config=yolov3-tiny.cfg --model=yolov3-tiny.weights --input=test.mp4 --width=416 --height=416 --classes=coco.names --scale=0.00392 --rgb --backend=2 --target=3

Where are:

--backend 0(auto), 1(Halide), 2(Intel-Engine), 3(OpenCV-impl)
--target 0(CPU), 1(OpenCL), 2(OpenCV-FP16), 3(VPU)