Peeling the Onion: Hierarchical Reduction of Data Redundancy for Efficient Vision Transformer Training

TL;DR

This paper introduces Tri-Level E-ViT, a hierarchical data redundancy reduction framework that accelerates Vision Transformer training by exploiting sparsity at multiple levels, often improving accuracy while reducing training time.

Contribution

It presents a novel end-to-end training method that reduces data redundancy across three hierarchical levels, enhancing efficiency without sacrificing accuracy.

Findings

Achieves up to 15.7% training speedup on ViT models.

Maintains or slightly improves Top-1 accuracy during acceleration.

Demonstrates the existence of significant data redundancy in ViT training.

Abstract

Vision transformers (ViTs) have recently obtained success in many applications, but their intensive computation and heavy memory usage at both training and inference time limit their generalization. Previous compression algorithms usually start from the pre-trained dense models and only focus on efficient inference, while time-consuming training is still unavoidable. In contrast, this paper points out that the million-scale training data is redundant, which is the fundamental reason for the tedious training. To address the issue, this paper aims to introduce sparsity into data and proposes an end-to-end efficient training framework from three sparse perspectives, dubbed Tri-Level E-ViT. Specifically, we leverage a hierarchical data redundancy reduction scheme, by exploring the sparsity under three levels: number of training examples in the dataset, number of patches (tokens) in each example, and number of connections between tokens that lie in attention weights. With extensive experiments, we demonstrate that our proposed technique can noticeably accelerate training for various ViT architectures while maintaining accuracy. Remarkably, under certain ratios, we are able to improve the ViT accuracy rather than compromising it. For example, we can achieve 15.2% speedup with 72.6% (+0.4) Top-1 accuracy on Deit-T, and 15.7% speedup with 79.9% (+0.1) Top-1 accuracy on Deit-S. This proves the existence of data redundancy in ViT.