Below you can find a outline of how to reproduce my solution for the "NFL 1st and Future - Impact Detection" competition.

ARCHIVE CONTENTS

lib : core source code

lib/scripts: python scripts to run data preparation, training, submitting etc. You needn't run them manually

lib/scripts/configs: various configuration files

models : trained models

checkpoints : provisional files saved during training

scripts : shell scripts to run data preparation, training, submitting. This scripts run corresponding python scripts with proper parameters. See entry_points.md for more info.

submissions : this is the place where the script ./scripts/predict.sh saves submissions

logs : logs directory

tmp : temporary directory

directory_structure.txt : contains directory structure of the project

entry_points.md : entry points to run data preparation, training, submitting

ModelSummary.odt : summary of my solution. It is also available on kaggle forum

README.md :

requirements.txt : requriments for pip

settings.json : main configuration file with paths to essential directories

HARDWARE: (The following specs were used to create the original solution)

Ubuntu 18.04.5 LTS

Intel(R) Core(TM) i7-8700K CPU @ 3.70GHz, 32 GB memory

2 x NVIDIA 1080 Ti

SOFTWARE (python packages are detailed separately in requirements.txt):

Python 3.8.5

CUDA 10.2

cuddn 7.6.5_0

nvidia drivers v440.118.02

If you use docker image like pytorch/pytorch you may need to install this software:

./scripts/install_other_soft.sh

After installation of packages from requirements.txt you need to install SlowFast library running script: ./scripts/install_slowfast.sh from the top directory.

DATA SETUP (assumes the Kaggle API is installed)

mkdir -p /mnt/SSDData/AF/
cd /mnt/SSDData/AF/
kaggle competitions download -c nfl-impact-detection
unzip nfl-impact-detection.zip

DATA PROCESSING

Add path to the unzipped files' directory into ./settings.json as RAW_DATA_DIR

Run ./scripts/prepare_data.sh. It takes half an hour to process the data and needs 60GB of free disk space in the directory TRAIN_PREPARED_DATA_PATH.

MODEL BUILD

1. very fast prediction
   • there is no very fast prediction in my solution
2. ordinary prediction
   • expect this to run for 4-6 hours for 15 pairs of videos
   • uses binary model files
3. retrain models
   • expect this to run about tho days
   • trains all models from scratch
   • follow this with (2) to produce entire solution from scratch

shell command to run each build is below:

1. None

2. ordinary prediction (uses pairs of raw videos from TEST_DATA_PATH and overwrites predictions in ./submissions directory)

./scripts/predict.sh

3. retrain models (overwrites models in ./checkpoints and ./models directory)

./scripts/train_models.sh