cs492_sbs

diet_for_kaist_cafeteria
if you have any questions, contact [email protected]
also pretrained_model might be deleted in my google drive, when long time after

I recommend you do these things in ubuntu 20.04 (my case)
##########################################

0. make conda environment
   conda create --name NAME python=3.8
   conda activate NAME
   pip install -r require_sbs.txt

1. download pretrained files ( my google drive links connected )
   python download.py

2. execute main file
   python CS492_final_sbs.py

3. execution guideline first of all prepare not tilted image of kaist cafeteria food.

'predict_calories' button is automatically executed
you can test the images by clicking buttons in order also.

########################################

For training yolo

########################################

0. to learn single images for food go to https://aihub.or.kr/aidata/30747 to download them. you should unzip and use json2txt_labe.py to change labeling format. type python json2txt_label.py [AI hub label directory] cs492_project_training/datasets/[Directory Name]/labels Also, image files should be moved to cs492_project_training/datasets/[Directory Name]/images

to download kaist food data python download_data.py and unzip them

1. to train use command

python train.py --data [DATA] --cfg [VERSION] --weights [PRETRAIN] --batch-size [SIZE]

commands that we used are following

for AI hub: python train.py --data food.yaml --cfg yolov5s --weights ''

for KAIST images: python train.py --data kama_final.yaml --weights [AIhub best.pt directory]

for plate detection: python train.py --data plates.yaml

you can refer detail instructions in https://github.com/ultralytics/yolov5

2. If you didn't add parameter for directory for results, it will be saved in ./runs/train/

   Best model is saved as "best.pt

3. KAIST food images are included and you can train them with kama_final.yaml

   AI hub images can be trained with food.yaml

!!We have some problems in pusing files. We tried to upload some important files for running our code. !!However, if there's some problem, please refer yolov5 github and use their files. !!Also, you can ask us for some problems.

########################################

For training segmentation

########################################
go to directory 'cs492_project/segmentation/'

1. you can make COCO image datasets[for single croped image] using 'labelme' and 'labelme2coco',
   1-1. i already give some of datasets for users.
   1-2. if you want to put your custom datasets, prepare COCO image dataset[for single croped image] into directory 'cs492_project/segmentation/images'
2. do not modify detectron2_train.py settings.
3. just execute
   python detectron2_train.py
4. final weight might be in cs492_project/segmentation/output/mode_final.pth
5. move it to 'cs492_project/' dierectory
6. Rename it 'model_final_best.pth'
   ########################################

this project thanks to yolov5, Detectron2 open source

its code was edited little bit by my self.