- clone repo
git clone https://github.com/grk717/automl-itmo.git
- pull data and weights
git lfs pull
- install requirements from requirements.txt
pip install -r requirements.txt
- run EDA notebook
icebergs_eda.ipynb
, it will create some data used in future - run LAMA notebook
icebergs_lama.ipynb
, it will create some data used in future - run
icebergs_manual.ipynb
, follow the instructions in the notebook
Link to the competition: https://www.kaggle.com/competitions/statoil-iceberg-classifier-challenge/
Metric for the competition is LogLoss
Made 4 submission, 3 of them are covered in this repo:
- LAMA with no EDA features (not covered) - 0.35363 on private data
- LAMA with EDA features - 0.33921 on private data
- LAMA with EDA features (added LGBM to blending, changed visual encoder backbone) - 0.32255 on private data
- YOLOv8 nano classifier - 0.26574 on private data