AI Drone Edu

< prerequisite >

conda deactivate

.bashrc edit

 

mkdir AI_Drone

cd AI_Drone

git clone https://github.com/ikelee77/inference_python.git

git clone https://github.com/AINukeHere/DroneFestival_U.git

( if pyqt5 is not installed, install it especially Dongmyung U.)

check

sudo apt-get install python3-pyqt5

install

sudo apt-get install python3-pyqt5

 

< training >

( cifar10 sample training )

cd ~/caffe

./data/cifar10/get_cifar10.sh

 -> downloard into ./data/cifar10 dir

create lmdb

./examples/cifar10/create_cifar10.sh

 -> generating lmdb in ./examples/cifar10 *.lmdb, *.lmdb, mean.binaryproto

(training)

edit : cifar10_full_solver.prototxt

 -> snapshot_format: BINARYPROTO

./examples/cifar10/train_quick.sh

( cifar10 sample inference )

needed files : *.caffemodel, *.prototxt

cp cifar10_full_iter_70000.caffemodel cifar.caffemodel

cp cifar10_full_train_test.prototxt cifar.prototxt

 

 

rename : cifar.caffemodel, cifar.prototxt

copy into ~/my_model

edit cifar.prototxt

name: "cifar"

delete

add layer(맨 위, 이름 바로 아래)

delete(매 뒤)

맨 뒤줄에 추가

layer {
  name: "prob"
  type: "Softmax"
  bottom: "ip1"
  top: "prob"
}

 

cd  /opt/intel/openvino/deployment_tools/model_optimizer/
sudo python3 mo/front/caffe/proto/generate_caffe_pb2.py --input_proto /home/intel/caffe/src/caffe/proto/caffe.proto

caffe_pb2.py  file is generated.

(optimize)

python3 mo.py --input_model ~/my_model/cifar.caffemodel --output_dir ~/my_model

 -> generated file into ~/my_model : cifar.bin, cifar.xml, cifar.mappin

<openvino 에러가 나면 필요한 모듈을 pip로 설치하고 다시 시도>

 

(inference test)

$   cd ~/sample
$   python3 rt_inference.py

 

---------

( image add )

image data conversion into png file

source /opt/intel/openvino/bin/setupvars.sh
 cd ~/caffe/examples/cifar10
 python3 cifar2png.py

 -> generated into ~/caffe/data/cifar10

image capture by cam

mkdir /home/intel/caffe/data/cifar10/capture
 python3 caphand.py

image downloard from kaggle (login)

https://www.kaggle.com/alishmanandhar/rock-scissor-paper/version/1

unzip and move to caffe/data/final

cd ~/caffe/examples/cifar10 
 python3 handmade.py

 -> converted image saved in caffe/data/cifar10 rock, scissors, paper (resized 32 x 32)

merge images kaggle and capture

createdb

edit createdb.py : add rock, scissors, paper into name array

python3 create.py ; test rate : 16

 -> generating lmdb in ./examples/cifar10 *.lmdb, *.lmdb

image_mean 파일 생성 

cd ~/caffe/
./build/tools/compute_image_mean -backend=lmdb examples/cifar10/cifar10_train_lmdb examples/cifar10/mean.binaryproto

 

< model optimizing >

copy model file into ~/my_model

and rename cifar.caffemodel

 

edit

edit cifar.prototxt : --> num_output change 10 or 13 or 16

 

Caffe용 prptotxt 환경 생성

cd  /opt/intel/openvino/deployment_tools/model_optimizer/
$   sudo python3 mo/front/caffe/proto/generate_caffe_pb2.py --input_proto /home/intel/caffe/src/caffe/proto/caffe.proto

최적화 진행

python3 mo.py --input_model ~/my_model/cifar.caffemodel --output_dir ~/my_model

  -> 3 files are generated

*.xml, *.bin, *.mapping

< inference >

코드 실행

~/my.model/*.xml, *.bin

cd ~/sample
 python3 rt_inference.py

drone : setting file

ObjectClassifier_Cifar10.py