Recommenders / Personalization

NVIDIA Merlin Accelerates Recommender Workflows with .4 Release

Relevant recommenders have the potential to impact millions of human decisions each day and build trust. Today, data scientists and machine learning engineers responsible for building relevant and impactful recommenders face challenges including slow pipelines, large embedding tables exceeding memory, and maintaining high throughput while maintaining low latency. These challenges are not inconsequential and can provide obstacles to frequently training, retraining, or deploying models into production. With this latest .4 release, NVIDIA Merlin delivers a new API and inference support that helps streamline the recommender workflow.

 Why Deep Learning for Recommenders?

Deep learning techniques enable machine learning engineers and data scientists to build fresh and relevant recommenders on large datasets at scale. Merlin democratizes building effective deep learning recommenders with open source components including: NVTabular, a preprocessing and feature engineering library; HugeCTR, a training framework created to handle large embedding tables; and Triton Inference Server which enables model deployment.    

Merlin Inference

With Merlin .4, both NVTabular and HugeCTR deepen inference support and integration with Triton Inference Server. As Triton Inference Server provides high performance throughput with low latency when deploying models, Merlin’s reinforces NVIDIA’s commitment to tackling common recommender challenges and accelerating workflows. 

Merlin API: Ease of Use

Merlin is currently in open beta and NVIDIA reaffirms commitment to streamlining recommender workflows by incorporating customer feedback into each release. The latest high level Merlin API makes it easier to define workflows and training pipelines.

Figure 1. Merlin NVTabular API
Figure 1-2: Merlin HugeCTR API

Download and Try All of Merlin Components

Building fresh and relevant recommenders at scale is a known challenge. Data scientists and machine learning engineers tackle these recommender challenges with a variety, or hybrid, of tools, techniques, and algorithms. Merlin is interoperable, designed to support machine learning engineers and scientists with preprocessing, feature engineering, training, and inference that helps provide relevant, impactful, and fresh recommenders at scale. If interested in accelerating recommender workflows and trying out an end-to-end deep learning recommender framework, visit the Merlin product home page or download from the NGC catalog.

Related resources

Discuss (0)

Tags