24-2 시각지능학습 코드 통합

How to Conduct a Peer Review for This Project

1. Clone the repository

git clone https://github.com/hy-vision-learning/jck-vision-int.git

2. Move to jck-vision-init

cd ./jck-vision-init

3. Install Required Packages

All required packages can be installed via requirements.txt.

pip install -r requirements.txt

Note

If an error related to package versions occurs during installation, remove the version information and try again.

4. Change the random seed

Open the change_randomseed.py file and change the random seed.

5. Run the code

We are ensembling WideResNet, PyramidNet, and DenseNet. Therefore, a total of four runs are required.

Please train the models by running ensemble-pyramidnet.ipynb, ensemble-densenet.ipynb, and ensemble-wideresnet.ipynb one at a time. Once all the training is complete, run ensemble-final.ipynb to output the final results.

Note

1. Please run only one .ipynb file at a time.
2. If you stop a running .ipynb file and need to retrain, restart the kernel before running it again.

6. Check the Results

The training time for each model will be displayed immediately after the training is completed.

The final training results will be output in the final file.

Final Results

seed	time	score
4943872	48h 43m	272.049

It is expected to take around 22 to 24 hours in practice.

main.py 사용법

argument 출력

python3 main.py --help

출력

  -h, --help            show this help message and exit
  -t TEST, --test TEST  테스트모드
  -pm MODEL_PATH, --model_path MODEL_PATH
                        모델 폴더 이름
  --amp AMP             amp 옵션
  -rs RANDOM_SEED, --random_seed RANDOM_SEED
                        학습 랜덤 시드. -1은 랜덤 시드를 고정하지 않음.
  -lf LOG_FILE, --log_file LOG_FILE
                        로그 파일 출력 여부. 0=false, 1=true
  -po PORT, --port PORT
  -m {custom,resnet18,resnet34,resnet50,resnet101,resnet152,resnext50,wide_resnet_16_4,wide_resnet_28_10_03,densenet121,densenet169,densenet201,densenet161,pyramidnet100_84,pyramidnet200_240,pyramidnet236_220,pyramidnet272_200,pyramidnet_custom}, --model {custom,resnet18,resnet34,resnet50,resnet101,resnet152,resnext50,wide_resnet_16_4,wide_resnet_28_10_03,densenet121,densenet169,densenet201,densenet161,pyramidnet100_84,pyramidnet200_240,pyramidnet236_220,pyramidnet272_200,pyramidnet_custom}
                        학습 모델
  -p PARALLEL, --parallel PARALLEL
                        멀티 gpu 사용 여부. 0=false, 1=true
  -op {sgd,adam,sam}, --optimizer {sgd,adam,sam}
                        옵티마이저
  -ls {none,lambda_lr,step_lr,cos_annealing,custom_annealing,one_cycle,cycle,on_plateau}, --lr_scheduler {none,lambda_lr,step_lr,cos_annealing,custom_annealing,one_cycle,cycle,on_plateau}
                        lr 스케쥴러
  -ds SPLIT_RATIO, --split_ratio SPLIT_RATIO
                        train/validation 분할 비율
  -am AUGMENTATION_MODE, --augmentation_mode AUGMENTATION_MODE
                        data augmentation mode
  -asp AUGMENT_SPLIT, --augment_split AUGMENT_SPLIT
                        augmentation 분할 비율
  -w NUM_WORKER, --num_worker NUM_WORKER
                        train/validation 분할 비율
  -b BATCH_SIZE, --batch_size BATCH_SIZE
                        학습 배치사이즈
  -mc MIX_STEP, --mix_step MIX_STEP
                        mix 적용시 몇 step마다 적용할지. 0은 모든 step에 적용.
  -mt {none,mixup,cutmix}, --mix_method {none,mixup,cutmix}
                        mix 방법
  -pd P_DEPTH, --p_depth P_DEPTH
                        pyramnidnet depth
  -pa P_ALPHA, --p_alpha P_ALPHA
                        pyramnidnet alpha
  -ps P_SHAKE, --p_shake P_SHAKE
                        pyramnidnet shake
  -e EPOCH, --epoch EPOCH
                        epoch
  -mlr MAX_LEARNING_RATE, --max_learning_rate MAX_LEARNING_RATE
                        optimizer/scheduler max learning rate 설정 (custom cos scheduler는 반대)
  -milr MIN_LEARNING_RATE, --min_learning_rate MIN_LEARNING_RATE
                        optimizer/scheduler min learning rate 설정 (custom cos scheduler는 반대)
  -wd WEIGHT_DECAY, --weight_decay WEIGHT_DECAY
                        optimizer weight decay 설정
  -gc GRADIENT_CLIP, --gradient_clip GRADIENT_CLIP
                        gradient clip 설정. -1은 비활성화
  -lsm LABEL_SMOOTHING, --label_smoothing LABEL_SMOOTHING
                        label smoothing 설정
  -es EARLY_STOPPING, --early_stopping EARLY_STOPPING
                        ealry stoppin epoch 지정. -1은 비활성화
  -ad ADAPTIVE, --adaptive ADAPTIVE
                        adaptive SAM 사용 여부
  -snt NESTEROV, --nesterov NESTEROV
                        nesterov sgd 사용 여부
  --rho RHO             SAM rho 파라미터
  -cm COS_MAX, --cos_max COS_MAX
                        cos annealing 주기
  -cp CUT_P, --cut_p CUT_P
                        cutmix 적용 확률
  -sm STEP_MILESTONE [STEP_MILESTONE ...], --step_milestone STEP_MILESTONE [STEP_MILESTONE ...]
                        step lr scheduler milestone