I implement face detection based on the idea of ​​yolo network in Kendryte K210. ps : this model only 500kb

you should install python package:

tensorflow 1.12 imgaug scikit-image numpy opencv-python matpltlib scipy tqdm

1. prepare dataset and make train list

   cd k210-face-detection
   wget http://tamaraberg.com/faceDataset/originalPics.tar.gz
   mkdir FDDB
   tar -zxvf originalPics.tar.gz -C FDDB
   wget http://vis-www.cs.umass.edu/fddb/FDDB-folds.tgz
   tar -zxvf FDDB-folds.tgz -C FDDB
   python3 tools/make_list.py --fddb_dir FDDB --ann_dir FDDB/FDDB-folds

   now your data fold have train.list

   NOTE: When generate the train.list I set the box width and height only half. if you want use the normal value to train, please remove the comment in tools/make_list.py line 47

2. train

   make train_pureconv ILR=0.001 MAXEP=20 IAA=false

   ILR : the init learning rate

   MAXEP : max epoch

   IAA : whether to use data augmenter

   NOTE: you can use CKPT:xxxxx to continue train

   example:

   make train_pureconv CKPT=log/20190216-152633 ILR=0.0005 MAXEP=20 IAA=true

   And you can use tensorboard --logdir log to look the record

make inference PB=Freeze_save.pb

1. freeze ckpt

   make freeze CKPT=log/20190216-154422 

   now your fold will have Freeze_save.pb

2. use kendryte-model-complier to complie pb file

   you can use my script (you should modify MODELCMP):

   make kmodel_convert PB=Freeze_save.pb MODELCMP=~/Documents/kendryte-model-compiler

   or refer to see the kendryte-model-compiler

1. copy weights array to code

   cp ~/Documents/kendryte-model-compiler/build/gencode_output.* K210_code/

2. compile the code

   you can refer to the documents k210 use in windows or k210 use in linux

   And you can find some useful article in my bolg.

3. down load program

   now you can use kflash.py down load the program

• [] add nms to avoid box overlapping

kendryte github

YAD2k