4th Place Solution - NFL Player Contact Detection

overview

This repository covers the training code for 2nd stage GBDT models in our pipeline.
The rest of the code can be found in the following repositories.

K_mat's NN training:
https://www.kaggle.com/code/kmat2019/nfl-training-sample-4thplace-kmatpart

Camaro's NN training:
https://github.com/bamps53/kaggle-nfl2022-4h-place-solution

As for details of our entire solution, Please check out this discussion.
https://www.kaggle.com/competitions/nfl-player-contact-detection/discussion/391761

2nd Stage GBDT models training

1. Preparation

# oof of camaro NN
kaggle datasets download -d bamps53/camaro-exp117
unzip camaro-exp117.zip -d ../input/camaro-exp117

kaggle datasets download -d bamps53/nfl-exp048
unzip nfl-exp048.zip -d ../input/nfl-exp048

# oof of kmat NN
kaggle datasets download -d kmat2019/mfl2cnnkmat0221
unzip mfl2cnnkmat0221.zip -d ../input/mfl2cnnkmat0221

# install
pip install -r requirements.txt

2. Training

Please execute notebook/train-gbdt.ipynb
When the execution is complete, the following 4 GBDT models are saved for the inference notebook.

Directory	Feature Set	Model	Corresponding Datasets
../input/nyanp-model-a-0227	K_mat(A)+Camaro	LightGBM	nyanpn/nyanp-model-a-0227
../input/nyanp-model-b-0227	K_mat(B)+Camaro	LightGBM	nyanpn/nyanp-model-b-0227
../input/nfl-kmat-only-2	K_mat(B)	LightGBM	nyanpn/nfl-kmat-only-2
../input/nfl-xgb-8030	K_mat(B)+Camaro	XGBoost	nyanpn/nfl-xgb-8030

3. Inference

Plase refer to this notebook.
Our best submission is made of above 4 GBDT models and 1 camaro 2nd stage model.
https://www.kaggle.com/code/bamps53/lb0796-exp184-185-lgb095?scriptVersionId=120623474

nyanp/nfl-player-contact-detection