A multi-entity Transformer for multi-agent spatiotemporal modeling.

Last update: Nov 15, 2022

Related tags

Overview

`baller2vec`

This is the repository for the paper:

Michael A. Alcorn and Anh Nguyen. baller2vec: A Multi-Entity Transformer For Multi-Agent Spatiotemporal Modeling. arXiv. 2021.


Left: the input for `baller2vec` at each time step t is an unordered set of feature vectors containing information about the identities and locations of NBA players on the court. Right: `baller2vec` generalizes the standard Transformer to the multi-entity setting by employing a novel self-attention mask tensor. The mask is then reshaped into a matrix for compatibility with typical Transformer implementations.

Left: the input for baller2vec at each time step t is an unordered set of feature vectors containing information about the identities and locations of NBA players on the court. Right: baller2vec generalizes the standard Transformer to the multi-entity setting by employing a novel self-attention mask tensor. The mask is then reshaped into a matrix for compatibility with typical Transformer implementations.


By exclusively learning to predict the trajectory of the ball, `baller2vec` was able to infer idiosyncratic player attributes.


Further, nearest neighbors in `baller2vec`'s embedding space are plausible doppelgängers. Credit for the images: Erik Drost, Keith Allison, Jose Garcia, Keith Allison, Verse Photography, and Joe Glorioso.


Additionally, several attention heads in `baller2vec` appear to perform different basketball-relevant functions, such as anticipating passes. Code to generate the GIF was adapted from @linouk23's NBA Player Movement's repository.


Here, a `baller2vec` model trained to simultaneously predict the trajectories of all the players on the court uses both the historical and current context to forecast the target player's trajectory at each time step. The left grid shows the target player's true trajectory at each time step while the right grid shows `baller2vec`'s forecast distribution. The blue-bordered center cell is the "stationary" trajectory.

Here, a baller2vec model trained to simultaneously predict the trajectories of all the players on the court uses both the historical and current context to forecast the target player's trajectory at each time step. The left grid shows the target player's true trajectory at each time step while the right grid shows baller2vec's forecast distribution. The blue-bordered center cell is the "stationary" trajectory.

Citation

If you use this code for your own research, please cite:

@article{alcorn2021baller2vec,
   title={baller2vec: A Multi-Entity Transformer For Multi-Agent Spatiotemporal Modeling},
   author={Alcorn, Michael A. and Nguyen, Anh},
   journal={arXiv preprint arXiv:1609.03675},
   year={2021}
}

Training `baller2vec`

Setting up `.basketball_profile`

After you've cloned the repository to your desired location, create a file called .basketball_profile in your home directory:

nano ~/.basketball_profile

and copy and paste in the contents of .basketball_profile, replacing each of the variable values with paths relevant to your environment. Next, add the following line to the end of your ~/.bashrc:

source ~/.basketball_profile

and either log out and log back in again or run:

source ~/.bashrc

You should now be able to copy and paste all of the commands in the various instructions sections. For example:

echo ${PROJECT_DIR}

should print the path you set for PROJECT_DIR in .basketball_profile.

Installing the necessary Python packages

cd ${PROJECT_DIR}
pip3 install --upgrade -r requirements.txt

Organizing the play-by-play and tracking data

Copy events.zip (which I acquired from here [mirror here] using https://downgit.github.io) to the DATA_DIR directory and unzip it:

mkdir -p ${DATA_DIR}
cp ${PROJECT_DIR}/events.zip ${DATA_DIR}
cd ${DATA_DIR}
unzip -q events.zip
rm events.zip

Descriptions for the various EVENTMSGTYPEs can be found here (mirror here).

Clone the tracking data from here (mirror here) to the DATA_DIR directory:

cd ${DATA_DIR}
git clone [email protected]:linouk23/NBA-Player-Movements.git

A description of the tracking data can be found here.

Generating the training data

cd ${PROJECT_DIR}
nohup python3 generate_game_numpy_arrays.py > data.log &

You can monitor its progress with:

top

or:

ls -U ${GAMES_DIR} | wc -l

There should be 1,262 NumPy arrays (corresponding to 631 X/y pairs) when finished.

Animating a sequence

If you don't have a display hooked up to your GPU server, you'll need to first clone the repository to your local machine and retrieve certain files from the remote server:

# From your local machine.
mkdir -p ~/scratch
cd ~/scratch

username=michael
server=gpu3.cse.eng.auburn.edu
data_dir=/home/michael/baller2vec_data
scp ${username}@${server}:${data_dir}/baller2vec_config.pydict .

games_dir=${data_dir}/games
gameid=0021500622

scp ${username}@${server}:${games_dir}/\{${gameid}_X.npy,${gameid}_y.npy\} .

You can then run this code in the Python interpreter from within the repository (make sure you source .basketball_profile first if running locally):

import os

from animator import Game
from settings import DATA_DIR, GAMES_DIR

gameid = "0021500622"
try:
    game = Game(DATA_DIR, GAMES_DIR, gameid)
except FileNotFoundError:
    home_dir = os.path.expanduser("~")
    DATA_DIR = f"{home_dir}/scratch"
    GAMES_DIR = f"{home_dir}/scratch"
    game = Game(DATA_DIR, GAMES_DIR, gameid)

# https://youtu.be/FRrh_WkyXko?t=109
start_period = 3
start_time = "1:55"
stop_period = 3
stop_time = "1:51"
game.show_seq(start_period, start_time, stop_period, stop_time)

to generate the following animation:

Running the training script

Run (or copy and paste) the following script, editing the variables as appropriate.

#!/usr/bin/env bash

# Experiment identifier. Output will be saved to ${EXPERIMENTS_DIR}/${JOB}.
JOB=$(date +%Y%m%d%H%M%S)

# Training options.
echo "train:" >> ${JOB}.yaml
task=ball_traj  # ball_traj, ball_loc, event, player_traj, score, or seq2seq.
echo "  task: ${task}" >> ${JOB}.yaml
echo "  min_playing_time: 0" >> ${JOB}.yaml  # 0/13314/39917/1.0e+6 --> 100%/75%/50%/0%.
echo "  train_valid_prop: 0.95" >> ${JOB}.yaml
echo "  train_prop: 0.95" >> ${JOB}.yaml
echo "  train_samples_per_epoch: 20000" >> ${JOB}.yaml
echo "  valid_samples: 1000" >> ${JOB}.yaml
echo "  workers: 10" >> ${JOB}.yaml
echo "  learning_rate: 1.0e-5" >> ${JOB}.yaml
if [[ ("$task" = "event") || ("$task" = "score") ]]
then
    prev_model=False
    echo "  prev_model: ${prev_model}" >> ${JOB}.yaml
    if [[ "$prev_model" != "False" ]]
    then
        echo "  patience: 5" >> ${JOB}.yaml
    fi
fi

# Dataset options.
echo "dataset:" >> ${JOB}.yaml
echo "  hz: 5" >> ${JOB}.yaml
echo "  secs: 4" >> ${JOB}.yaml
echo "  player_traj_n: 11" >> ${JOB}.yaml
echo "  max_player_move: 4.5" >> ${JOB}.yaml
echo "  ball_traj_n: 19" >> ${JOB}.yaml
echo "  max_ball_move: 8.5" >> ${JOB}.yaml
echo "  n_players: 10" >> ${JOB}.yaml
echo "  next_score_change_time_max: 35" >> ${JOB}.yaml
echo "  n_time_to_next_score_change: 36" >> ${JOB}.yaml
echo "  n_ball_loc_x: 95" >> ${JOB}.yaml
echo "  n_ball_loc_y: 51" >> ${JOB}.yaml
echo "  ball_future_secs: 2" >> ${JOB}.yaml

# Model options.
echo "model:" >> ${JOB}.yaml
echo "  embedding_dim: 20" >> ${JOB}.yaml
echo "  sigmoid: none" >> ${JOB}.yaml
echo "  mlp_layers: [128, 256, 512]" >> ${JOB}.yaml
echo "  nhead: 8" >> ${JOB}.yaml
echo "  dim_feedforward: 2048" >> ${JOB}.yaml
echo "  num_layers: 6" >> ${JOB}.yaml
echo "  dropout: 0.0" >> ${JOB}.yaml
if [[ "$task" != "seq2seq" ]]
then
    echo "  use_cls: False" >> ${JOB}.yaml
    echo "  embed_before_mlp: True" >> ${JOB}.yaml
fi

# Save experiment settings.
mkdir -p ${EXPERIMENTS_DIR}/${JOB}
mv ${JOB}.yaml ${EXPERIMENTS_DIR}/${JOB}/

# Start training the model.
gpu=0
cd ${PROJECT_DIR}
nohup python3 train_baller2vec.py ${JOB} ${gpu} > ${EXPERIMENTS_DIR}/${JOB}/train.log &

A multi-entity Transformer for multi-agent spatiotemporal modeling.

Related tags

Overview

`baller2vec`

Citation

Training `baller2vec`

Setting up `.basketball_profile`

Installing the necessary Python packages

Organizing the play-by-play and tracking data

Generating the training data

Animating a sequence

Running the training script

Owner

Michael A. Alcorn

Classical OCR DCNN reproduction based on PaddlePaddle framework.

Code for T-Few from "Few-Shot Parameter-Efficient Fine-Tuning is Better and Cheaper than In-Context Learning"

Trading environnement for RL agents, backtesting and training.

A simple API wrapper for Discord interactions.

Codebase for ECCV18 "The Sound of Pixels"

Cleaned up code for DSTC 10: SIMMC 2.0 track: subtask 2: multimodal coreference resolution

A cool little repl-based simulation written in Python

Implementation of Sequence Generative Adversarial Nets with Policy Gradient

Ganilla - Official Pytorch implementation of GANILLA

Addon and nodes for working with structural biology and molecular data in Blender.

Explanatory Learning: Beyond Empiricism in Neural Networks

ilpyt: imitation learning library with modular, baseline implementations in Pytorch

Auto White-Balance Correction for Mixed-Illuminant Scenes

Self-supervised Deep LiDAR Odometry for Robotic Applications

[ICCV 2021] Deep Hough Voting for Robust Global Registration

This is a repository with the code for the ACL 2019 paper

LiDAR R-CNN: An Efficient and Universal 3D Object Detector

The FIRST GANs-based omics-to-omics translation framework

A new framework, collaborative cascade prediction based on graph neural networks (CCasGNN) to jointly utilize the structural characteristics, sequence features, and user profiles.

Manage the availability of workspaces within Frappe/ ERPNext (sidebar) based on user-roles

A multi-entity Transformer for multi-agent spatiotemporal modeling.

Related tags

Overview

baller2vec

Citation

Training baller2vec

Setting up .basketball_profile

Installing the necessary Python packages

Organizing the play-by-play and tracking data

Generating the training data

Animating a sequence

Running the training script

Owner

Michael A. Alcorn

Classical OCR DCNN reproduction based on PaddlePaddle framework.

Code for T-Few from "Few-Shot Parameter-Efficient Fine-Tuning is Better and Cheaper than In-Context Learning"

Trading environnement for RL agents, backtesting and training.

A simple API wrapper for Discord interactions.

Codebase for ECCV18 "The Sound of Pixels"

Cleaned up code for DSTC 10: SIMMC 2.0 track: subtask 2: multimodal coreference resolution

A cool little repl-based simulation written in Python

Implementation of Sequence Generative Adversarial Nets with Policy Gradient

Ganilla - Official Pytorch implementation of GANILLA

Addon and nodes for working with structural biology and molecular data in Blender.

Explanatory Learning: Beyond Empiricism in Neural Networks

ilpyt: imitation learning library with modular, baseline implementations in Pytorch

Auto White-Balance Correction for Mixed-Illuminant Scenes

Self-supervised Deep LiDAR Odometry for Robotic Applications

[ICCV 2021] Deep Hough Voting for Robust Global Registration

This is a repository with the code for the ACL 2019 paper

LiDAR R-CNN: An Efficient and Universal 3D Object Detector

The FIRST GANs-based omics-to-omics translation framework

A new framework, collaborative cascade prediction based on graph neural networks (CCasGNN) to jointly utilize the structural characteristics, sequence features, and user profiles.

Manage the availability of workspaces within Frappe/ ERPNext (sidebar) based on user-roles

`baller2vec`

Training `baller2vec`

Setting up `.basketball_profile`