OpenVINO Training Extensions - BitVehicle

[junxnone]

Reference

• [ ] Train SSD detection model

  Setup the Environment

  . venv/bin/activate

  Downloads

• Download BitVehicle Dataset
• MobileNet v2 0.35 256x256

Pre-processing dataset

• Unpack dataset in the directory images

  cd data/bitvehicle
  unrar x BITVehicle_Dataset.rar  
  mv BITVehicle_Dataset images

• Downscale images to increase training speed

  python ./tools/downscale_images.py -target_size 512 ./data/bitvehicle/images 

  modify the config.py to specified the GPU.

  vi vlp/config.py

Training

python3 train.py vlp/config.py

The model.ckpt will save in the training_toolbox/ssd_detector/vlp/model/model.ckpt.

Evaluation

python3 eval.py vlp/config.py

Inference

python3 infer.py --json --input=../../data/bitvehicle/bitvehicle_test.json  --show vlp/config.py

This will show the inference with the test image.

Export OpenVINO format model

 python3 export.py vlp/config.py /opt/intel/computer_vision_sdk/deployment_tools/model_optimizer/mo.py

$ tree vlp/model/ie_model/
vlp/model/ie_model/
├── checkpoint
├── graph.bin
├── graph.ckpt.data-00000-of-00001
├── graph.ckpt.index
├── graph.mapping
├── graph.pb
├── graph.pb.frozen
├── graph.pbtxt
├── graph.tfmo.json
└── graph.xml

Inference with OpenVINO

python3 object_detection_demo_ssd_async.py -m vlp/model/ie_model/graph.xml -l /opt/intel/computer_vision_sdk/inference_engine/lib/ubuntu_16.04/intel64/libcpu_extension_avx2.so -i xxx.mp4

[junxnone]

79