Portal de Periódicos da CAPES

Qtorch+: Next Generation Arithmetic for Pytorch Machine Learning

2022; Springer Science+Business Media; Linguagem: Inglês

10.1007/978-3-031-09779-9_3

1611-3349

Nhut-Minh Ho, Himeshi De Silva, John L. Gustafson, Weng‐Fai Wong,

Digital Filter Design and Implementation

This paper presents Qtorch+, a tool which enables next generation number formats on Pytorch, a widely popular high-level Deep Learning framework. With hand-crafted GPU accelerated kernels for processing novel number formats, Qtorch+ allows developers and researchers to freely experiment with their choice of cutting-edge number formats for Deep Neural Network (DNN) training and inference. Qtorch+ works seamlessly with Pytorch, one of the most versatile DNN frameworks, with little added effort. At the current stage of development, we not only support the novel posit number format, but also any other arbitrary set of points in the real number domain. Training and inference results show that a vanilla 8-bit format would suffice for training, while a format with 6 bits or less would suffice to run accurate inference for various networks ranging from image classification to natural language processing and generative adversarial networks. Furthermore, the support for arbitrary number sets can contribute towards designing more efficient number formats for inference in the near future. Qtorch+ and tutorials are available on GitHub ( https://github.com/minhhn2910/QPyTorch ).