A Deep Learning Dataloader with Shared Data Preparation

Authors: jian xie, Jingwei Xu, Guochang Wang, Yuan Yao, Zenan Li, Chun Cao, Hanghang Tong

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the proposed JOADER , showing a greater versatility and superiority of training speed improvement (up to 200% on Res Net18) without affecting the accuracy.
Researcher Affiliation Academia 1Nanjing University 2University of Illinois Urbana-Champaign
Pseudocode No The paper describes algorithms and data structures through textual descriptions and diagrams, but no formal pseudocode or algorithm blocks were found.
Open Source Code No The paper states that a prototype named JOADER was implemented, but it does not provide any concrete access information for its source code.
Open Datasets Yes In this section, we evaluate JOADER on Image Net with the family of Res Net models.
Dataset Splits No The paper evaluates on ImageNet but does not explicitly provide specific training/validation/test dataset splits (percentages, counts, or detailed methodology for splitting).
Hardware Specification Yes The experiments were conducted on a GPU server with two Intel Xeon Gold 5118 CPUs @ 2.30GHz (24 physical cores and 48 threads), 500GB RAM, and 6 TITAN RTX GPUs. The server ran Ubuntu 18.04 with GNU/Linux kernel 4.15.0. The disk is Symbios Logic Mega RAID SAS-3 3316 of 1GB/s read speed.
Software Dependencies Yes The evaluated models are the basic models with their default settings in torchvision [1], and trained on top of the Py Torch 1.6.0 DL framework. The server ran Ubuntu 18.04 with GNU/Linux kernel 4.15.0.
Experiment Setup Yes The evaluated models are the basic models with their default settings in torchvision [1], and trained on top of the Py Torch 1.6.0 DL framework. In this experiment, we start training multiple Res Net18 models at the same time but with different hyper-parameters.