Level2-ObjectDetection Competetion

🌟CV-08조🌟 MakeZenerator 팀 김태양, 김혜지, 신호준, 성주희, 임서현, 정소윤

Project Structure

${PROJECT}
 ┃  
 ┣ Data_EDA
 ┃ ┗ visualizer.ipynb  
 ┃ ┗ visualizer_save.py 
 ┃
 ┣ dataset.py
 ┣ east_dataset.py
 ┣ file_rename.py
 ┣ json_convert.py
 ┣ detect.py
 ┣ loss.py
 ┣ model.py
 ┣ deteval.py
 ┣ train.py
 ┣ inference.py
 ┃
 ┣ .github
 ┣ .gitignore  
 ┣ .gitmessage.txt  
 ┣ .pre-commit-config.yaml  
 ┗ README.md  

• Data_EDA : This folder contains..
• FOR dataset : dataset.py, east_dataset.py, file_rename.py, json_convert.py, detect.py
• FOR Models : loss.py, model.py, deteval.py
• FOR Tools : train.py, inference.py
• README.md
• requirements.txt : contains the necessary packages to be installed

Code Structure Description

Data EDA

• Use MaskSplitByProfileDataset
• Downsampling
• Stratified Kfold

Model

• Ensemble Soft Voting
• Learn additional Fine Tuning based on the public pretrained model EfficientNet + ConvNext + ConvNext(Stratified Kfold)

Getting Started

Setting up Vitual Enviornment

1. Initialize and update the server

   su -
   source .bashrc
   

2. Create and Activate a virtual environment in the project directory

   conda create -n env python=3.8
   conda activate env
   

3. To deactivate and exit the virtual environment, simply run:

   deactivate
   

??? Install Requirements

To Insall the necessary packages liksted in requirements.txt, run the following command while your virtual environment is activated:

pip install -r requirements.txt

Usage

Description of all arguments

Training

To train the model with your custom dataset, set the appropriate directories for the training images and model saving, then run the training script.

• single model

python train.py --data_dir /path/to/images --model_dir /path/to/model --model MODEL_NAME

• single multiple model

python train_single_multiple.py --data_dir /path/to/images --model_dir /path/to/model --model MODEL_NAME

Inference

For generating predictions with a trained model, provide directories for evaluation data, the trained model, and output, then run the inference script.

• single model

python inference.py --data_dir /path/to/images --model_dir /path/to/model --output_dir /path/to/model --model MODEL_NAME

• single multiple model

python inference.py --data_dir /path/to/images --model_dir /path/to/model --output_dir /path/to/model --model_mode single_multiple --model MODEL_NAME

Ensemble

• ensemble (hard voting)

python hard_voting.py --file_dir ./csv --csv1 file1.csv --csv2 file2.csv --csv3 file3.csv

• ensemble (soft voting)

python soft_voting.py --models MODEL_NAME1 MODEL_NAME2 MODEL_NAME3 --model_dir ./checkpoint --model_files file1.pth file2.pth file3.pth --data_dir ./data/eval

Arguments

Model

• ConvNext Tiny
• EfficientNet v2 small
• ViT16

Dataloader

• oversampling with data augmentation