TRAINING MATHEMATICAL EXPRESSION RECOGNITION MODEL WITH MULTIPLE GPUs

Abstract

This paper reports the findings of an trial in which multiple GPUs were used to train mathematical expression recognition in Deep Learning. The different learning strategies were investigated using a Convolutional Neural Network (CNN). The dataset used for the training was IM2LATEX-100K, provided by WYGIWYS containing 103,556 different LaTeX expressions extracted from more than 60,000 articles from arXiv. Pytorch Lingting library with a convolutional sequence modeling network with ConvMath was employed on a computer system with 4 GPUs. The findinds showed an increased training speed with multiple GPUs.