Instance Segmentation on Distributed Deep Learning Big Data Cluster

Résumé

Abstract This paper presents workflow in deep learning inference models in a distributed manner. This is often required in production environments where real-time predictions need to be made on a massive scale., optimization approaches were applied to the instance segmentation YOLACT model to improve the inference throughput. , and the experiments show that the execution time taken by the OpenVino (Open Visual Inference and Neural Network Optimization) model with 16 FP is the least compared to ONNX (Open Neural Network Exchange) and the original PyTorch model. ONNX and OpenVINO provide a set of tools for optimizing deep learning models for deployment on a range of edge devices such as smartphones, laptops, tablets, IoT devices, sensors, and other devices that collect, process, and transmit data between the end user and the cloud or datacenter. The model optimizer in these frameworks plays a critical role in preparing deployment models, by reducing their size and increasing their performance.YOLACT in OpenVino format we run it on a big data cluster using the BigDL framework to apply inference on spark stand-alone and yarn clusters. where BigDL is a distributed deep learning library for Apache Spark that provides a high-level programming interface for defining and training deep neural networks. It supports a wide range of deep learning models and is designed to be scalable, making it ideal for large-scale deep learning applications. In distributed deep learning reference, the input data is partitioned and sent to multiple machines in parallel for processing. Each machine runs its portion and produces a partial output, which is then combined with the outputs of other machines to produce a final result. Spark stand alone perform better depending on the obtaining results under different conditions whereas significant speedups can be achieved by increasing the number of executors across the cluster.

Citer ce document

Accès au document

Texte intégral en lecture en ligne, réservé aux abonnés SPHAERO et aux membres de l'institution. Se connecter

Auteur(s)

Statistiques

Consultations : 1

Téléchargements : 0