yolo : https://github.com/AlexeyAB/darknet.git
yolo to tflite : https://github.com/hunglc007/tensorflow-yolov4-tflite.git
라벨링
https://github.com/tzutalin/labelImg.git

sudo apt-get install pyqt5-dev-tools voc -> yolo 포맷 변경: https://github.com/ssaru/convert2Yolo.git

준비

darknet , yolo_mark (동영상 자를때), labelImg 설치
git clone git@github.com:cheolgyu/gamebot-yolo.git
darknet 실행파일 gamebot-yolo에 넣고
git clone git@github.com:cheolgyu/gamebot-dataset.git
gamebot-dataset 이동해서 gamebot-dataset/ds/폴더 원드라이드랑 업로드 동기화 시켜 동영상 다운받기
gamebot-yolo docker는 사용안하고 우분투에서 실행시킴.
사진없고 동영상만 있으면 yolo_mark로 프레임 짜르고
사진가지고 라벨링을하는데 labelImg로 하기
라벨링 어느정도 완료하면 라벨링된것을 훈련용과 검증용 으로 분류해야됨.
분류스크립트는 https://github.com/cheolgyu/gamebot-dataset/blob/master/yolo/label_after2.py 참고하고
obj.data 랑 obj.names랑 backup폴더 확인하고
학습하기 처음-이어서로 학습하면되고 그래프가안나오면 opnecv가 설치제대로안해서 그런거니깐 보고싶으면 설치 ㄱㄱ
중간중간 훈련된것 확인하는방법
훈련이 어느정된후 tflite로 바꾸는방법은
tensorflow-yolov4-tflite 폴더에 converet_dockerfile 빌드후에 실행시키고
거기서 weights to tensorflow to tflite 참고해서 실행시키면 됨. 2번실행해야됨.tensorflow(pb) 바꾸고 그담 그걸 tflite로 만들고 이렇게 2번실행.

프로젝트들

v4
gotgl: 왕좌의게임:윈터이즈커밍 갱신

fivestars 파이브스타즈

yolov4-tiny.cfg
classes= 8
0. 쾌시작
1. 쾌정보-확인
2. 전투시작
3. 월드맵
4. 건너뛰기
5. 쾌스트보상-확인
6. 렙업확인
7. 컨텐츠오픈 확인

illusionc 일루전커넥트 yolov4-tiny.cfg 0. 쾌시작
1. 도전
2. 알림
3. 승리
4. 스킵
5. 패배
6. 획득보상 classes=12
sk2 세븐나이츠2

    project5

        인식대상기준
            0 0.956291 0.050000 0.058007 0.088406   --홈
            1 0.933415 0.906522 0.066176 0.140580   --절전풀.방치o
            2 0.097631 0.057246 0.090686 0.097101   --전체풀
            3 0.769608 0.775362 0.225490 0.136232   --분해좌
            4 0.645016 0.675362 0.100490 0.081159   --일반
            5 0.743056 0.674638 0.097222 0.073913   --고급
            6 0.903595 0.771014 0.181373 0.168116   --분해우
            7 0.230801 0.547826 0.100490 0.168116   --금화위치.분해결과2.분해선택
            7 0.184641 0.547101 0.086601 0.157971   --금화위치.분해결과3.분해선택

            8 0.524101 0.500000 0.776961 0.515942   --분해팝업창
            9 0.613971 0.663043 0.247549 0.155072   --분해팝업창.확인
            10 0.428922 0.479710 0.104575 0.171014  --금화위치.분해결과2
            10 0.361928 0.484783 0.104575 0.169565  --금화위치.분해결과3
            11 0.498775 0.221014 0.178922 0.120290  --분해결과.텍스트
            12 0.928102 0.072500 0.119361 0.085000  --스킵
            13 0.613715 0.699846 0.204861 0.121914  --스킵확인
            14 0.387587 0.699074 0.203993 0.111111  --스킵취소
            15 0.879229 0.790833 0.121241 0.208333  --스마.off
            16 0.879229 0.790833 0.121241 0.208333  --스마.on
            17 0.500000 0.498457 0.913194 0.651235  --팀교체.전체
            18 0.486979 0.650463 0.477431 0.331790  --팀교체.중간
            19 0.601562 0.715278 0.197917 0.158951  --팀교체.진행
            20 0.293837 0.401235 0.556424 0.709877  --도움.전체
            21 0.296441 0.109568 0.459201 0.120370  --도움.중간
            22 0.492622 0.105710 0.070312 0.128086  --도움.닫기
            23 0.522059 0.492972 0.727376 0.638554  --전투패배팝업

opncv 설치

opncv install scripts/install_opnecv4.sh

라벨 변경

convert2Yolo의 예대로 폴더구조 변경해야됨.

라벨링

https://github.com/AlexeyAB/darknet#datasets

build

docker build -t gb-yolo:latest .

docker run --name gb-yolo -d -it -v ~/workspace/gb-yolo:/workspace/gb-yolo -p 8888:8888 gb-yolo:latest scripts/install_OpenCV4.sh ~/workspace/darknet/build.sh docker cp gb-yolo:/workspace/darknet/darknet ./darknet

cfg

https://github.com/AlexeyAB/darknet#how-to-train-to-detect-your-custom-objects
max_batches to (classes*2000
change [filters=255] to filters=(classes + 5)x3 in the 3 [convolutional] before each [yolo]

학습하기

처음

./darknet detector train workspace/baram/p1/obj.data cfg/baram-p1-3l.cfg yolov4-tiny.conv.29 -map ./darknet detector train workspace/sk2/p7/obj.data cfg/sk2_p7.cfg yolov4-tiny.conv.29 -map

이어서

./darknet detector train workspace/baram/p1/obj.data cfg/baram-p1-3l.cfg workspace/baram/p1/backup/baram-p1-3l_last.weights -map ./darknet detector train workspace/sk2/project_1/obj.data cfg/sk2_2_yolov4-tiny-3l.cfg workspace/sk2/project_1/backup/sk2_2_yolov4-tiny-3l_last.weights -map -show_imgs ./darknet detector train workspace/illusionc/p1/obj.data cfg/illusionc_1.cfg workspace/illusionc/p1/backup/illusionc_1_last.weights -map ./darknet detector map workspace/illusionc/p1/obj.data cfg/illusionc_1.cfg workspace/illusionc/p1/backup/illusionc_1_last.weights

test

./darknet detector demo workspace/baram/p1/obj.data cfg/baram_p1.cfg workspace/baram/p1/backup/baram-p1-3l_last.weights -ext_output /home/cheolgyu/다운로드/baram_003.mp4 -thresh 0.6 ./darknet detector demo workspace/sk2/project_1/obj.data cfg/sk2_2_yolov4-tiny-3l.cfg workspace/sk2/project_1/backup/sk2_2_yolov4-tiny-3l_best.weights -ext_output /home/cheolgyu/다운로드/sk2_0021.mp4

./darknet detector test workspace/sk2/project_1/obj.data cfg/sk2_1.cfg workspace/sk2/project_1/backup/sk2_1_last.weights -thresh 0.25 ./darknet detector test workspace/sk2/project_1/obj.data cfg/sk2_2_yolov4-tiny-3l.cfg workspace/sk2/project_1/backup/sk2_2_yolov4-tiny-3l_best.weights -thresh 0.25

./darknet_cpu detector test workspace/v4/project_3/obj.data cfg/yolov4-tiny-3l-v4-project_3.cfg workspace/v4/project_3/backup/yolov4-tiny-3l-v4-project_3_best.weights -thresh 0.01

./darknet detector demo workspace/illusionc/p1/obj.data cfg/illusionc_1.cfg workspace/illusionc/p1/backup/illusionc_1_last.weights -ext_output /home/cheolgyu/다운로드/gotgl_video_11.mp4

./darknet detector map workspace/illusionc/p1/obj.data cfg/illusionc_1.cfg workspace/illusionc/p1/backup/illusionc_1_last.weights

사진

./darknet detector test 오브젝트.데이터 cfg파일 무게 -thresh 0.25 ./darknet detector test /home/cheolgyu/workspace/gb-yolo/workspace/blackdesertm/project_1/obj.data /home/cheolgyu/workspace/gb-yolo/cfg/yolov4-tiny-baram.cfg /home/cheolgyu/workspace/gb-yolo/workspace/blackdesertm/project_1/backup/yolov4-tiny-3l_baram_crop_best.weights -thresh 0.25

./darknet detector test /home/cheolgyu/workspace/gb-yolo/workspace/blackdesertm/project_1/obj.data /home/cheolgyu/workspace/gb-yolo/cfg/yolov4-tiny-3l_baram_crop.cfg /home/cheolgyu/workspace/gb-yolo/workspace/blackdesertm/project_1/backup/yolov4-tiny-3l_baram_crop_best.weights -thresh 0.25

동영상

./darknet detector demo 오브젝트.데이터S cfg파일 무게 -ext_output 동영상

weights to tensorflow to tflite -container run

update classes /data/classes

512

python save_model.py --weights data/baram-p1-3l_last.weights --output ./checkpoints/baram-p1-3l-640 --input_size 640 --model yolov4 --framework tflite --tiny

python convert_tflite.py --weights ./checkpoints/sk2-p7-448 --output ./checkpoints/sk2-p7-448/sk2-p7-448.tflite

yolov4 quantize int8

python convert_tflite.py --weights ./checkpoints/sk2-p7-448 --output ./checkpoints/sk2-p7-448-int8.tflite --quantize_mode int8 --dataset ./coco_dataset/coco/val207.txt

yolov4 quantize float16

python convert_tflite.py --weights ./checkpoints/baram-p1-3l-640 --output ./checkpoints/baram-p1-3l-640-fp16.tflite --quantize_mode float16 python detect.py --weights ./checkpoints/baram-p1-3l-640-fp16.tflite --size 640 --model yolov4 --image ./baram_0022_00000000.jpg --framework tflite

yolov4 quantize float16

python convert_tflite.py --weights ./checkpoints/kor-p1-448 --output ./checkpoints/kor-p1-448-fp16.tflite --quantize_mode float16 python detect.py --weights ./checkpoints/kor-p1-448-fp16.tflite --size 448 --model yolov4 --image ./kor_0013_00000646.jpg --framework tflite

python detect.py --weights ./checkpoints/sk2-p7-448-fp16.tflite --size 448 --model yolov4 --image ./sk2_0022_00000066.jpg --framework tflite

python detect.py --weights ./checkpoints/sk2_p6-448/sk2_p6-448.tflite --size 448 --model yolov4 --image ./sk2_0070_00000000.jpg --framework tflite

640

python save_model.py --weights data/sk2_p5_yolov4-tiny-3l_best.weights --output ./checkpoints/sk2_p5-640 --input_size 640 --model yolov4 --framework tflite --tiny

python convert_tflite.py --weights ./checkpoints/sk2_p5-640 --output ./checkpoints/sk2_p5-640/sk2_p5-640.tflite

python detect.py --weights ./checkpoints/sk2_p4-640/sk2_p4-640.tflite --size 640 --model yolov4 --image ./sk2_test_img/sk2_0020_00000493.jpg --framework tflite

416

python save_model.py --weights data/illusionc/illusionc_1_last.weights --output ./checkpoints/illusionc_1_last-416 --input_size 416 --model yolov4 --framework tflite --tiny

python convert_tflite.py --weights ./checkpoints/illusionc_1_last-416 --output ./checkpoints/illusionc_1_last-416/illusionc_1_last-416.tflite

python detect.py --weights ./checkpoints/illusionc_1_last-416 --size 416 --model yolov4 --image ./illusionc_video_8_00000078.jpg --tiny

python detect.py --weights ./checkpoints/illusionc_1_last-416/illusionc_1_last-416.tflite --size 416 --model yolov4 --image ./illusionc_v6_00000063.jpg --framework tflite

832

python save_model.py --weights data/gotgl/gotgl_14_last.weights --output ./checkpoints/gotgl_14_last-832 --input_size 832 --model yolov4 --framework tflite --tiny

python convert_tflite.py --weights ./checkpoints/gotgl_14_last-832 --output ./checkpoints/gotgl_14_last-832/gotgl_14_last-832.tflite

python detect.py --weights ./checkpoints/gotgl_14_last-832 --size 832 --model yolov4 --image ./gotgl_video_8_00000078.jpg --tiny

python detect.py --weights ./checkpoints/gotgl_14_last-832/gotgl_14_last-832.tflite --size 832 --model yolov4 --image ./gotgl_video_8_00000078.jpg --framework tflite

960

python save_model.py --weights data/gotgl/gotgl_15_last.weights --output ./checkpoints/gotgl_15_last-960 --input_size 960 --model yolov4 --framework tflite --tiny

python convert_tflite.py --weights ./checkpoints/gotgl_15_last-960 --output ./checkpoints/gotgl_15_last-960/gotgl_15_last-960.tflite

python detect.py --weights ./checkpoints/gotgl_15_last-960 --size 960 --model yolov4 --image ./gotgl_video_8_00000078.jpg --tiny

python detect.py --weights ./checkpoints/gotgl_15_last-960/gotgl_15_last-960.tflite --size 960 --model yolov4 --image ./gotgl_video_5_00000292.jpg --framework tflite gotgl_video_8_00000078 gotgl_video_5_00000292

1280

python save_model.py --weights data/gotgl/gotgl_14_last.weights --output ./checkpoints/gotgl_14_last-1280 --input_size 1280 --model yolov4 --framework tflite --tiny

python convert_tflite.py --weights ./checkpoints/gotgl_14_last-1280 --output ./checkpoints/gotgl_14_last-1280/gotgl_14_last-1280.tflite

python detect.py --weights ./checkpoints/gotgl_14_last-1280 --size 1280 --model yolov4 --image ./gotgl_video_8_00000078.jpg --tiny

python detect.py --weights ./checkpoints/gotgl_14_last-1280/gotgl_14_last-1280.tflite --size 1280 --model yolov4 --image ./gotgl_video_8_00000078.jpg --framework tflite

weights to tensorflow to tflite ==>2

960x480

python save_model.py --weights data/gotgl/gotgl_11_last.weights --output ./checkpoints/gotgl_11-960x480 --input_size 960x480 --model yolov4 --framework tflite --tiny python convert_tflite.py --weights ./checkpoints/gotgl_11-960x480 --output ./checkpoints/gotgl_11-960x480/gotgl_11-960x480.tflite python detect.py --weights ./checkpoints/gotgl_11-960x480/gotgl_11-960x480.tflite --size 960x480 --model yolov4 --image ./gotgl_video_8_00000078.jpg --framework tflite

832x832

python save_model.py --weights data/gotgl/gotgl_11_last.weights --output ./checkpoints/gotgl_11-832x832 --input_size 832x832 --model yolov4 --framework tflite --tiny python convert_tflite.py --weights ./checkpoints/gotgl_11-832x832 --output ./checkpoints/gotgl_11-832x832/gotgl_11-832x832.tflite python detect.py --weights ./checkpoints/gotgl_11-832x832/gotgl_11-832x832.tflite --size 832x832 --model yolov4 --image ./gotgl_video_8_00000078.jpg --framework tflite

608x608

python save_model.py --weights data/gotgl/gotgl_11_last.weights --output ./checkpoints/gotgl_11-608x608 --input_size 608x608 --model yolov4 --framework tflite --tiny python convert_tflite.py --weights ./checkpoints/gotgl_11-608x608 --output ./checkpoints/gotgl_11-608x608/gotgl_11-608x608.tflite python detect.py --weights ./checkpoints/gotgl_11-608x608/gotgl_11-608x608.tflite --size 608x608 --model yolov4 --image ./gotgl_video_8_00000078.jpg --framework tflite

416x416

python save_model.py --weights data/gotgl/gotgl_11_last.weights --output ./checkpoints/gotgl_11-416x416 --input_size 416x416 --model yolov4 --framework tflite --tiny python convert_tflite.py --weights ./checkpoints/gotgl_11-416x416 --output ./checkpoints/gotgl_11-416x416/gotgl_11-416x416.tflite python detect.py --weights ./checkpoints/gotgl_11-416x416/gotgl_11-416x416.tflite --size 416x416 --model yolov4 --image ./gotgl_video_8_00000078.jpg --framework tflite

Run demo tflite model

python detect.py --weights ./checkpoints/yolov4-tiny-3l-baram_crop_best.tflite --size 832 --model yolov4 --image ./data/1597947045328.jpg --framework tflite

python detect.py --weights ./checkpoints/yolov4-tiny-3l-baram_crop_best --size 832 --model yolov4 --image ./data/7_쾌버튼1-회색-수락하기.jpg --tiny

cheolgyu/gamebot-yolo

준비

프로젝트들

opncv 설치

라벨 변경

라벨링

build

cfg

학습하기

처음

이어서

test

사진

동영상

weights to tensorflow to tflite -container run

update classes /data/classes

512

yolov4 quantize int8

yolov4 quantize float16

yolov4 quantize float16

640

416

832

960

1280

weights to tensorflow to tflite ==>2

960x480

832x832

608x608

416x416

Run demo tflite model