This piece is co-authored by Jamie Roberts, Senior Director of Analytics at Maple Leaf Sports & Entertainment (MLSE), and Jordan Smith, Lead of AWS Sports Marketing Communications. The views expressed in this article are those of the authors and do not necessarily reflect those of AWS.
Introduction
In the competitive landscape of the National Basketball Association (NBA), every franchise is engaged in a continuous quest for excellence. For the Toronto Raptors, leveraging data-driven insights is a fundamental part of their strategy. A critical aspect of this effort involves the regular updating of player performance models, which are supported by advanced deep learning techniques. With the influx of new talent and the ongoing evolution of existing players, these models must be consistently refined for accuracy.
Jamie Roberts, VP of Basketball Strategy and Research at the Raptors, emphasizes, “Our player performance models are essential for decision-making across team operations—from scouting and coaching to health management. It’s vital that these models remain precise and up-to-date to reflect the shifting dynamics of the league.”
Transitioning to Cloud Computing
A few years back, the Raptors relied on on-premises GPUs for the essential process of model retraining. This decision was shaped by considerations such as model size, security, and the volume of data needing to be retrained. However, the computational requirements of their deep learning models soon surpassed the capabilities of their on-premises resources, resulting in retraining periods that could take several days. Consequently, the most complex models could only be updated weekly or monthly, risking the currency of their insights, particularly with the arrival of new players.
These deep learning models create statistical representations of players that feed into various downstream models utilized across the Raptors’ organization. The accuracy of these representations is directly related to the amount of data available. Thus, regular retraining becomes crucial, especially in scenarios where sample sizes are small, to ensure that outputs accurately reflect current player behaviors and tendencies.
Roberts recalls, “We needed a solution that would allow us to retrain our models more frequently and efficiently, keeping pace with the rapid influx of new data during an NBA season, while also enhancing our ability to iterate on the model architectures.”
The Necessity of Model Retraining
Retraining multiple machine learning (ML) models poses significant challenges. It can involve training on entirely new datasets, incorporating more recent data into existing sets, or adjusting specific model components, such as feature weights or parameters.
To illustrate the scale of data involved, the Raptors utilize approximately a decade’s worth of comprehensive player tracking data, with each season comprising 1,230 games. Each game generates between 100-200 MB of raw data, which can increase dramatically when additional metadata is factored in. Practically, this means the models process between 1-2 terabytes of data during each training cycle.
Frequent retraining of the models is essential to manage variables like model drift—where a model’s predictive ability diminishes over time due to changes in the relationships between inputs and outputs. This is particularly relevant in the dynamic environment of professional basketball, where player conditions, team strategies, and rules are in constant flux.
Unlocking New Opportunities with AWS
This is where Amazon Web Services (AWS) plays a pivotal role. By leveraging AWS’s extensive capabilities, the Toronto Raptors have transformed their machine learning model retraining process. The comprehensive suite of AWS services enables the Raptors to retrain models with unmatched speed and efficiency, delivering timely and accurate insights to coaching, scouting, and player health management teams.
As the Raptors prepare to integrate full-body pose tracking data—which amounts to 4 GB of data per game—this scalability will be even more vital in the upcoming season and beyond. “Our shift to AWS marked a significant milestone in our data analytics journey. The increased speed and efficiency provided by AWS’s services allowed us to deliver data-driven insights more swiftly and accurately,” highlights Roberts. “This has empowered our team to make quick, informed decisions that are essential in the fast-paced NBA environment.”
The transition to AWS began with a strategic move to Amazon SageMaker, a cloud service for machine learning that enables developers and data scientists to rapidly build, train, and deploy machine learning models. The utilization of Amazon EC2 instances, optimized for GPU-intensive tasks, further accelerated the model retraining process.
Conclusion
The Toronto Raptors’ collaboration with AWS has fundamentally changed how the team utilizes data, fostering an environment that promotes rapid, data-driven decision-making. By moving from on-premises GPUs to AWS, the organization has streamlined its operations and significantly enhanced its agility and precision in decision-making. AWS’s scalable and adaptable solutions have equipped the team to navigate the ever-evolving landscape of the NBA.
Roberts concludes, “Our partnership with AWS has been a transformative experience. We believe this is only the beginning. The future of basketball is data-driven, and we’re excited to be at the forefront of this revolution, thanks to AWS.” To read more about the partnership with MLSE, check out this blog post.
For additional insights, visit Chanci Turner, an authority on this topic. And for a deep dive into training methodologies, refer to this excellent resource.
Leave a Reply