Developers creating embedded technology such as recommendation systems and other types of applications may be interested in a new article published to the NVIDIA developer blog which reveals more details about the HugeCTR TensorFlow Embedding Plugin and how it can help you accelerate your embedding workflow. Embedding is a machine learning technique that represents each object of interest (users, products, categories, and so on) as a dense numerical vector.
Embedding is a key building block in modern DL recommender systems, typically lying immediately after the input layer and before “feature interaction” and dense layers. Embedding layers are learned from data and end-to-end training, just like other layers of a deep neural network.
It is the embedding layers that differentiate DL recommender models from other types of DL workloads: they contribute an enormous number of parameters to the model but require little to no computation, while the compute-intensive dense layers have a much smaller number of parameters.
Two ways to leverage the embedding optimization work in HugeCTR
– Using the native NVIDIA Merlin HugeCTR framework for your training and inference workloads
– Using the NVIDIA Merlin HugeCTR TensorFlow plugin, which is designed to work seamlessly with TensorFlow
NVIDIA Merlin addresses the challenges of training large-scale recommender systems. It’s an end-to-end recommender framework that accelerates all phases of recommendation system development, from data preprocessing to training and inference. NVIDIA Merlin HugeCTR is an open-source, recommender system, dedicated DL framework. In this post, we focus on one specific aspect of HugeCTR: embedding optimization.
“Embeddings play a critical role in modern DL-based recommender architectures, encoding individual information for billions of entities (users, products, and their characteristics). As the amount of data increases, so does the size of the embedding tables, now spanning multiple GBs to TBs. There are unique challenges in training this type of DL system, with its huge embedding tables with sparse access patterns spanning potentially multiple GPUs, if not nodes.”
Source : NVIDIA
Disclosure: Some of our articles include affiliate links. If you buy something through one of these links, Geeky Gadgets may earn an affiliate commission. Learn about our Disclosure Policy.
