YOLOv3 implementation with Tensorflow on Android
This project contains an example of YoloV3 implementation on Android, the YoloV3 model was implemented through the library
org.tensorflow:tensorflow-android.
Below is a list of steps taken to convert the YoloV3 model from darkflow to tensorflow for Android (command launched on Ubuntu inside Anaconda):
- clone DW2TF repository from here https://github.com/jinyu121/DW2TF to local folder
- ex: /home/user/projects/
- Download (or train) YoloV3 model and weights in darknet format (.cfg and .weights)
- Launch DW2TF conversion as mentioned on the github page of DW2TF: https://github.com/jinyu121/DW2TF:
*
python3 main.py \ --cfg 'data/yolov3-tiny.cfg' \ --weights 'data/yolov3-tiny.weights' \ --output 'data/' \ --prefix 'yolov3-tiny/' \ --gpu 0 - launch freeze_graph to have a single bp graph file:
freeze_graph \
--input_graph yolov3-tiny.pb \
--input_checkpoint yolov3-tiny.ckpt \
--input_binary=true \
--output_graph=ultimate_yolov3.bp \
--output_node_names=yolov3-tiny/convolutional10/BiasAdd
NOTE: For older version of Yolo you can use darkflow tool https://github.com/thtrieu/darkflow, here an example after clone the repository:
./flow --model ../data/yolov2-tiny.cfg --load ../data/yolov2-tiny.weights --savepb
For more detail about Yolo look at offical page https://pjreddie.com/darknet/yolo/