Convert the weights of YOLO v3 object detector into tensorflow lite format. It can be served for tensorflow serving as well.

docker build -t tflite .
docker run -it -v /home/peace195/tensorflow-lite-yolo-v3:/root/ tflite

1. Download COCO class names file: wget https://raw.githubusercontent.com/pjreddie/darknet/master/data/coco.names
2. Download binary file with desired weights:
   • Full weights: wget https://pjreddie.com/media/files/yolov3.weights
   • Tiny weights: wget https://pjreddie.com/media/files/yolov3-tiny.weights
   • SPP weights: wget https://pjreddie.com/media/files/yolov3-spp.weights
3. Convert .weights to .pb saved_model python ./convert_weights_pb.py (this can be used for tensorflow serving)
4. Convert .pb to .tflite tflite_convert --saved_model_dir=saved_model/ --output_file yolo_v3.tflite --saved_model_signature_key='predict'

Optional Flags

convert_weights_pb.py:

--class_names
    Path to the class names file
--weights_file
    Path to the desired weights file    
--data_format
    `NCHW` (gpu only) or `NHWC`
--tiny
    Use yolov3-tiny
--spp
    Use yolov3-spp
--output_graph
    Location to write the output .pb graph

Contact me if you have any issues: [email protected] / Binh Do