CV_classification

Classification Pipeline in Computer Vision (Pytorch)

Environments

Directory

bash CV_classification ├── datasets │ ├── __init__.py │ ├── augmentation.py │ └── factory.py ├── models │ ├── __init__.py │ ├── resnet.py │ └── loss.py ├── log.py ├── main.py ├── train.py ├── run.sh ├── requirements.txt ├── README.md └── LICENSE

Pipeline

1. Set seed
2. Make directory to save results
3. Build model
4. Build dataset with augmentations
   • Train dataset
   • Validation dataset (optional)
   • Test dataset
5. Make dataLoader
6. Define optimizer (model parameters)
7. Define loss function
8. Training model
   • Checkpoint model using evaluation on validation dataset
   • Log training history using logging or wandb in save folder
9. Testing model

Run

run.sh

```bash dataname=$1 numclasses=$2 optlist='SGD Adam' lrlist='0.1 0.01 0.001' auglist='default weak strong' bs_list='16 64 256'

for bs in $bslist do for opt in $optlist do for lr in $lrlist do for aug in $auglist do # use scheduler echo "bs: $bs, opt: $opt, lr: $lr, aug: $aug, usesched: True" EXPNAME="bs$bs-opt$opt-lr$lr-aug$aug-use_sched"

            if [ -d "$EXP_NAME" ]
            then
                echo "$EXP_NAME is exist"
            else
                python main.py \
                    --exp-name $EXP_NAME \
                    --dataname $dataname \
                    --num-classes $num_classes \
                    --opt-name $opt \
                    --aug-name $aug \
                    --batch-size $bs \
                    --lr $lr \
                    --use_scheduler \
                    --epochs 50
            fi

            # not use scheduler
            echo "bs: $bs, opt: $opt, lr: $lr, aug: $aug, use_sched: False"
            EXP_NAME="bs_$bs-opt_$opt-lr_$lr-aug_$aug"

            if [ -d "$EXP_NAME" ]
            then
                echo "$EXP_NAME is exist"
            else
                python main.py \
                    --exp-name $EXP_NAME \
                    --dataname $dataname \
                    --num-classes $num_classes \
                    --opt-name $opt \
                    --aug-name $aug \
                    --batch-size $bs \
                    --lr $lr \
                    --epochs 50
            fi
        done
    done
done

done ```

example

bash bash run.sh CIFAR10 10

Config Parameters

Wandb 관련 설정

|Argument|Description|Default|Possible value| |---|---|---|---| |usewandb|Wandb 사용 여부|True|True,False| |usecm|Confusion metrix 사용 여부|True|True,False| |entity|Wandb 엔티티 명|"connect-cv-04"|---| |projectname|Wandb 프로젝트 명|"Imageclassificationmask"|---| |expname|실험명|"exp"|---| |expnum|실험 번호|0|---| |username|실험자|"my_name"|"KDH","KJY","HJH","KDK"|

실험 관련 설정

|Argument|Description|Default|Possible value| |---|---|---|---| |datadir|input 경로|'../input|---| |trainfile|train csv 이름|"train.csv"|---| |validfile|valid csv 이름|"valid.csv"|---| |transform|Transform 목록|['resize','randomrotation', 'totensor', 'normalize']|---| |seed|Random seed|223|---| |modelname|Modelnames|“CustomModel”|---| |modelparam|Modelnames|{pretrained : True, backbone : "resnet18"}|---| |numclasses|Class 개수|18|---| |batchsize|Batch size|32|---| |optname|Optimizer|"Adam"|"Adam"| |loss|loss 종류|"crossentropy"|"crossentropy","focalloss","f1loss","bceloss","mseloss"| |lossparam|loss parm|"미정"|---| |lr|learning rate|5e-6|---| |lrsheduler|"Learning rate scheduler|"StepLR"|"StepLR","ReduceLROnPlateau"| |lrshedulerparam|Lr scheduler parameter|"미정"|---| |weightdecay|Weight Decay|5e-4|---| |epochs|epoch|100|---| |savedir|모델 저장 위치|"./checkpoint"|---| |gradaccumsteps|---|1|---| |mixed_precision|---|"fp16"|---| |patience|Early Stopping|100|---|

Contributors ✨

Thanks goes to these wonderful people (emoji key):

김대희 💬 🚧 💻 👀 🤔

김동규 💻 🚧 💬 🔬 🔧

김준영 💻 🚇 🤔 📆🔬

황준하 💻 🤔 💬 🎨 📖

This project follows the all-contributors specification. Contributions of any kind welcome!