action-recognition

A Video Vision Transformer (ViViT) model for detecting incidences of contact between between NFL players in play-by-play video footage

Summary

• Containerized project using Docker to recreate project env (including OpenCV, PyTorch, hugginface, etc...)
• Uploaded data and Docker image to AWS S3 and ECR, respectively
• Finetuned pretrained ViViT model (from hugginface) on AWS EC2 GPU-enabled instance
• Model based on ViViT: A Video Vision Transformer

Setup

Assuming you have docker installed:

1. Pull data from: https://www.kaggle.com/competitions/nfl-player-contact-detection/data

2. Store data in project directory

    ├── ActionRecognition           <- Project directory
    │    ├── data                   <- Data directory
    │         ├── ...               <- Data files
   

3. Build docker image

$cd path/to/ActionRecognition
$docker build -t action-rec:latest .

4. Run docker container

$docker run -v $(pwd)/data:/usr/src/data -v $(pwd)/models:/usr/src/models -it action-rec:latest sh

5. Within container: Run inference

$python inference.py

Project Organization

├── README.md          <- Top-level README
|
├── models             <- Trained and serialized models
│     ├── ...          
│
├── requirements.txt   <- Requirements file for reproducing the analysis environment
|
├── .dockerignore      <- Files to omit when building image
│
├── Dockerfile         <- Instructions for building docker image
│
├── config.yaml        <- Model configurations/hyperparameters
|
├── dataset.py         <- Functions and class for building dataset for model training
|
├── train.py           <- Model training
|
├── inference.py       <- Model inference/testing