{"title":"bs-scheduler: A Batch Size Scheduler library compatible with PyTorch DataLoaders","authors":"George Stoica, Mihaela Elena Breabăn","doi":"10.1016/j.softx.2025.102162","DOIUrl":null,"url":null,"abstract":"<div><div>Deep learning models involve computationally intensive training experiments. Increasing the batch size improves the training speed and hardware efficiency by enabling deep neural networks to ingest and process more data in parallel. Inspired by learning rate adaptation policies that yield good results, methods that gradually adjust the batch size have been developed. These methods enhance hardware efficiency without compromising generalization performance. Despite their potential, such methods have not gained widespread popularity or adoption: unlike widely used learning rate policies, for which there is built-in support in most of the deep learning frameworks, the use of batch size adaptation policies requires custom implementations. We introduce an open-source package that implements batch size adaptation policies, which can be seamlessly integrated into deep learning training pipelines. This facilitates more efficient experimentation and accelerates research workflows.</div></div>","PeriodicalId":21905,"journal":{"name":"SoftwareX","volume":"30 ","pages":"Article 102162"},"PeriodicalIF":2.4000,"publicationDate":"2025-04-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"SoftwareX","FirstCategoryId":"94","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S2352711025001293","RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"COMPUTER SCIENCE, SOFTWARE ENGINEERING","Score":null,"Total":0}
引用次数: 0
Abstract
Deep learning models involve computationally intensive training experiments. Increasing the batch size improves the training speed and hardware efficiency by enabling deep neural networks to ingest and process more data in parallel. Inspired by learning rate adaptation policies that yield good results, methods that gradually adjust the batch size have been developed. These methods enhance hardware efficiency without compromising generalization performance. Despite their potential, such methods have not gained widespread popularity or adoption: unlike widely used learning rate policies, for which there is built-in support in most of the deep learning frameworks, the use of batch size adaptation policies requires custom implementations. We introduce an open-source package that implements batch size adaptation policies, which can be seamlessly integrated into deep learning training pipelines. This facilitates more efficient experimentation and accelerates research workflows.
期刊介绍:
SoftwareX aims to acknowledge the impact of software on today''s research practice, and on new scientific discoveries in almost all research domains. SoftwareX also aims to stress the importance of the software developers who are, in part, responsible for this impact. To this end, SoftwareX aims to support publication of research software in such a way that: The software is given a stamp of scientific relevance, and provided with a peer-reviewed recognition of scientific impact; The software developers are given the credits they deserve; The software is citable, allowing traditional metrics of scientific excellence to apply; The academic career paths of software developers are supported rather than hindered; The software is publicly available for inspection, validation, and re-use. Above all, SoftwareX aims to inform researchers about software applications, tools and libraries with a (proven) potential to impact the process of scientific discovery in various domains. The journal is multidisciplinary and accepts submissions from within and across subject domains such as those represented within the broad thematic areas below: Mathematical and Physical Sciences; Environmental Sciences; Medical and Biological Sciences; Humanities, Arts and Social Sciences. Originating from these broad thematic areas, the journal also welcomes submissions of software that works in cross cutting thematic areas, such as citizen science, cybersecurity, digital economy, energy, global resource stewardship, health and wellbeing, etcetera. SoftwareX specifically aims to accept submissions representing domain-independent software that may impact more than one research domain.