1st and Future - Player Contact Detection
Below you can find an outline of how to reproduce my 1st place solution for the 1st and Future - Player Contact Detection
.
1.INSTALLATION
- Ubuntu 18.04.5 LTS
- CUDA 11.2
- Python 3.7.5
- Training PC: 1x RTX3090 (or any GPU with at least 24Gb VRAM), 32GB RAM, at least 5TB of disk space.
- python packages are detailed separately in requirements.txt
$ conda create -n envs python=3.7.5
$ conda activate envs
$ pip install -r requirements.txt
$ pip install torch==1.10.1+cu113 torchvision==0.11.2+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html
$ pip install -U openmim
$ mim install mmcv-full
2.DATA
- Download dataset and extract to
data/
folder. - Download pretrained mmaction2 weight
wget -P pretrained/ "https://download.openmmlab.com/mmaction/recognition/csn/vmz/vmz_ircsn_ig65m_pretrained_r50_32x2x1_58e_kinetics400_rgb_20210617-86d33018.pth"
3. PREPROCESSING
- Create folds and cache helmet boxes and tracking data to dictionary.
$ python create_folds.py
$ python cache_metadata.py
- Train xgb preprocessing to filter easy negative sample.
$ cd tree/
$ python pre_g_xgb.py
$ python pre_p_xgb.py
- Generate cache for CNN model and save to disk.
$ cd cnn/
$ python create_cache11_G.py
$ python create_cache15_G.py
$ python create_cache15.py
- The project folder now looks like
├── data
│ ├── train
│ ├── test
│ ├── train_player_tracking.csv
│ ├── train_folds.csv
│ ├── ...
├── tree
│ ├── pp_g_xgb.py
│ ├── pp_p_xgb.py
│ ├── pre_g_xgb
│ ├── ...
├── cnn
│ ├── configs
│ │ ├── r50ir_csn_c11_m1_d2_G_all.py
│ │ ├── r50ir_csn_c15_m1_d2_all.py
│ │ ├── ...
│ ├── models
│ │ ├── model_csn1.py
│ │ ├── resnet3d_csn.py
│ ├── dataset
│ │ ├── dataset_3d_3ch_v2.py
│ ├── cache
│ │ ├── cache11_G
│ │ ├── cache15_G
│ │ ├── cache15
│ ├── pretrained/
│ ├── train.py
│ ├── inference.py
│ ├── create_cache11_G.py
│ ├── create_cache15_G.py
│ ├── create_cache15.py
│ ├── train.sh
│ ├── ...
├── cache_metadata.py
├── create_folds.py
├── readme.md
├── requirements.txt
4.TRAINING
- Preprocessing step must be run before this step.
- To train and validate all the models, run the following command inside the
cnn/
folder.
$ ./train.sh
- Folds 0,1,2,3,4 are used for validation and train a xgb post processing model.
- Folds 5,6,7,8,9 are trained with all data and used in submission. PP model used fold 5,6,7,8; PG model used fold 6,7,8 from 2 configs (6 checkpoints).
- Train xgb post processing.
$ cd tree/
$ python pp_g_xgb.py
$ python pp_p_xgb.py
5.INFERENCE
- Since this is a code competition, the inference code is available in kaggle notebook.
- https://www.kaggle.com/code/nvnnghia/nfl3-ensemble-w-pp-sub?scriptVersionId=121109062