SCI_Kostat2022

2022 통계데이터 인공지능 활용대회 GIST SCI LAB

Team Members

서동건(Team Leader)	김주영	김주연
GIST EECS Bachelor student	GIST IIT Combined MS/PhD student	GIST IIT MS student
Modeling, Tuning	EDA, Data cleaning	EDA, ML-based approach

😀Directory

kostat
├── SCI_Kostat2022
│   ├── README.md
│   ├── create_kfold.py
│   ├── dataset.py
│   ├── inference.py
│   ├── model.py
│   ├── preprocess.py
│   ├── requirements.txt
│   ├── train_MD.py
│   ├── train_WC.py
│   └── utils.py
└── input
    ├── 1. 실습용자료.txt
    ├── 2. 모델개발용자료.txt
    ├── 답안 작성용 파일.csv
    └── 한국표준산업분류(10차)_국문.xlsx

😀Environment setting

Default python version == 3.9.10

pip install -r requirements.txt

😀Train model(WC)

Using Default AutoModelForSequenceClassification from Huggingface transformers

python train_WC.py --kfold 5

😀Train model(MD)

Using Custom model which Multi-Dropout is applied (model is implemented in model.py)

python train_MD.py -=kfold 5

😀Inference

Choose best CV checkpoint model for each fold
All hyperparameters used to get results below are described in code

CV accuracy for each model

	WC	MD
fold0	93.032	92.963
fold1	92.965	92.954
fold2	93.015	93.033
fold3	92.954	92.991
fold4	92.918	92.954

## you can edit this checkpoint list depended on your result
model_checkpoint = [f'../best_model/roberta_large_WC_fold{fold}' for fold in range(0,2)]
model_checkpoint1 = [f'../best_model/roberta_large_WC_MD_fold{fold}' for fold in range(2,5)]
model_checkpoint.extend(model_checkpoint1)
inference(model_checkpoint) 

python inference.py