DENSO Automates ML Model Development for Driving Support Using Amazon SageMaker, Reduces Work Time by 55%–66%

DENSO Corporation (DENSO), a manufacturer specializing in automotive components, needs to keep up with the fast evolution of image sensors, which use machine learning (ML) to support safe driving. The company develops image sensors for advanced driver-assistance systems (ADAS), which help drivers with functions such as parking and changing lanes. To develop the necessary ML models for ADAS image recognition, DENSO had built GPU clusters in its on-premises environment. However, multiple ML engineers shared limited GPU resources, which impacted productivity—especially during the busy period before a new product release. Because DENSO operated and maintained its ML solution in house, the engineers also spent significant time managing servers and file systems and wanted to focus more on product innovation. Finally, DENSO needed to achieve a higher level of product traceability to satisfy international standards for transparency.

To address these challenges, DENSO used Amazon Web Services (AWS) and implemented Amazon SageMaker, which helps data scientists and developers to prepare, build, train, and deploy high-quality ML models quickly. “By adopting Amazon SageMaker, we were able to accelerate the creation of ADAS image recognition models by reducing the data acquisition, model development, learning, and evaluation time,” says Kensuke Yokoi, general manager at DENSO. “These time savings helped our engineers focus on the features that differentiate their offering, leaving the undifferentiated heavy lifting to AWS.”

Headquartered in Japan, DENSO has 200 subsidiaries and nearly 171,000 employees throughout the world. The company develops and manufactures automobile technology, systems, and products for uses such as advanced safety and autonomous driving (AD). The AD & ADAS Technology Division 1 at DENSO is responsible for developing AD and ADAS technology, including image recognition sensors that distinguish vehicles, people, lane markings, road signs, and more. “In this field, which has evolved significantly, the complexity and number of models has increased due to an increasing number of new scenarios and functions,” says Ryoma Niihara, project assistant manager at DENSO. DENSO needed to find a solution that would empower its engineers to keep up with these changes.

After comparing various on-premises and cloud-based solutions, DENSO decided to modernize its ML solution and replace its GPU clusters using Amazon SageMaker so that developers could more quickly create and use ML models. Using the Cloud Economics Center, which helps customers understand and optimize the value of AWS, DENSO calculated its total cost of ownership and discovered that it would save about 20 percent using Amazon SageMaker compared to other solutions. In addition, DENSO found Amazon SageMaker to be user friendly and appreciated the support available from AWS. “AWS gave us thorough support and an abundance of advice, from defining the requirements to studying the architecture,” says Shinji Niwa of the cloud service research and development division at DENSO. “We were new to using AWS for ML, and the support from AWS was very reliable.”

DENSO began building its ML solution in July 2020 using serverless services on AWS. It used AWS Step Functions alongside Amazon SageMaker to create a continuous integration / continuous delivery pipeline for its ML models. AWS Step Functions is a low-code visual workflow service used to orchestrate AWS services, automate business processes, and build serverless applications. DENSO used the AWS Step Functions Data Science Software Development Kit, which helps customers create multistep ML workflows in Python, to create workflows for preprocessing, postprocessing, and evaluating its ML models—making it simpler to implement quick changes and improve the models’ accuracy. By being particular about the architecture, it was simple for ML engineers to develop the environment based on a prototype.

In DENSO’s solution, the data to be used in ML models is consolidated in Amazon Simple Storage Service (Amazon S3), an object storage service that offers industry-leading scalability, data availability, security, and performance. DENSO’s ML engineers then forward the data to Amazon FSx for Lustre, a fully managed service that provides cost-effective, high-performance, scalable storage for compute workloads—thereby reducing the time it takes to process each model development job. Finally, DENSO acquires traceability data using Amazon DocumentDB (with MongoDB compatibility), a database service that is purpose built for JavaScript Object Notation data management at scale.

The AD & ADAS Technology Division 1 was ready to begin using its ML solution by January 2021. Because each GPU instance needed for ML is allocated using the continuous integration / continuous delivery pipeline, the team finds it easy to optimize its resources and manage costs. The data engineers at DENSO have reduced their hours spent on data management by 55 percent, and the ML engineers have reduced time spent on repeat work by 66 percent. By parallelizing ML jobs using Amazon SageMaker, the team has reduced learning time from 3 days to 3 hours. Soon after implementation, the division received praise from its artificial intelligence and ML engineers, who appreciated being able to automatically achieve traceability, perform repeat tasks by workflow, and manage big data processing with greater ease. “It’s a significant achievement,” says Yokoi. “The managed services helped us concentrate engineers’ resources on the work that adds the most value.”

Now that DENSO saves time and labor using its new ML solution, the company plans to transfer all remaining ML model deployment environments to Amazon SageMaker by around January 2022. The company is considering using Amazon SageMaker Ground Truth, a fully managed data labeling service, to streamline the annotation work required to tag its image data. DENSO might also expand its use of ML in the cloud for other sensor types—such as light detection and ranging, which is a remote sensing method that uses lasers to measure ranges.

“The practice of shifting to the cloud will keep accelerating in the artificial intelligence and ML field,” says Yokoi. “I’m confident that AWS will continue to give us support as we continue adding functions.”