/Classification_of_oil_condition

Primary LanguageJupyter Notebook

건설기계 오일 상태 분류 AI 경진대회

건설장비에서 작동오일의 상태를 실시간으로 모니터링하기 위한 오일 상태 판단 모델 개발 (정상, 이상의 이진분류)
Competition Link

  • 주최: 현대제뉴인 [Link]
  • 후원: AWS
  • 주관: Dacon
  • Private 12th, Score 0.577764

Structure

Train/Test data and sample submission file must be placed under dataset folder.

repo.
  |——dataset
        |——train.csv
        |——test.csv
        |——sample_submission.csv
        |——data_info.xlsx
  |——models
        |——teacher
            |——Teacher_model1.pickle
            |—— ...
            |——Teacher_model5.pickle
        |——student
            |——Student_model1.pickle
            |—— ...
            |——Student_model5.pickle
  |——result
        |——submission_CatBoost.csv
  |——ipynotebook
        |——학습코드.ipynb
  |——requirements.txt
  |——teacher_train.py
  |——student_train.py
  |——inference.py

Development Environment

Our Training Resource

Windows 10
11th Gen Intel(R) Core(TM) i7-11850H
NVIDIA RTX A2000 Laptop GPU

Or you can execute in Web based Development Services

If you want to run it with Colab or AWS Sagemaker, run the ipynb code below. Also, It includes an EDA and results analysis process. It will output the same result.

  • Google Colab
  • AWS Sagemaker Studio Lab
repo.
  |——ipynotebook
        |——학습코드.ipynb

Training Code Run Solution

0. Dependency

the details of packages version are listed in requirements.txt

> python --version
Python 3.9.15
> pip install -r requirements.txt
> pip list
...
numpy         1.23.5
pandas        1.5.2
scikit-learn  1.2.0
catboost      1.1.1
optuna        3.0.4
...

Because the number of variables in the learning environment and the inference environment are different, we used the knowledge distillation method. So, We go through 2 trainning process to obtain Teacher model and Student Model

1. Teacher Model Training

  • learning method: CatBoostClassifier
  • 5 StratifiedKFold ensemble
    • So, 5 model weights are saved after training
  • includes hyper-parameter tuning preocess with Optuna
> python teacher_train.py

Result

models
      |——teacher
            |——Teacher_model1.pickle
            |—— ...
            |——Teacher_model5.pickle

2. Student Model Training

  • train with the teacher's output as the label
  • learning method: CatBoostRegressor
  • 5 StratifiedKFold ensemble
    • So, 5 model weights are saved after training
  • includes hyper-parameter tuning preocess with Optuna
> python student_train.py

Result

models
      |——student
            |——Student_model1.pickle
            |—— ...
            |——Student_model5.pickle

3. Inference

  • 5 fold ensemble (soft-voting) inference
  • metric: F1 score
  • thershold: 0.15
python inference.py

Result

check submission_CatBoost.csv

result
      |——submission_CatBoost.csv

Download Best Model

You can download our best model weight Here GoogleDrive.

Inference With Best Model

Best model must be unzipped under models folder

1. Unzip Best Model

unzip and move best model to models directory

2. Inference

python inferecne.py

3. Result

check submission_CatBoost.csv

models  # (it is best model)
    |——teacher
          |——Teacher_model1.pickle
          |—— ...
          |——Teacher_model5.pickle
    |——student
          |——Student_model1.pickle
          |—— ...
          |——Student_model5.pickle
result  # (it is best result)
    |——submission_CatBoost.csv
inference.py