Domain aware post training quantization for vision transformers in deployment

Li Wang

doi:10.1016/j.patcog.2025.112182

Domain aware post training quantization for vision transformers in deployment

2025-07-24

0

PRE

AI

DOI：10.1016/j.patcog.2025.112182

Original

Original request for help

Save

Abstract

En 中文

• Introduced a novel Domain Aware Post-Training Quantization (DAQuant) approach tailored for Vision Transformers (ViTs) to address performance degradation due to model compression and domain shift. • Developed a distribution-guided activation smoothing and scaling method, alongside Learnable Activation Clipping (LAC), to mitigate outliers and improve quantization accuracy under extreme low-bit conditions. • Proposed an effective domain alignment strategy that enhances generalization ability on the target domain while preserving optimization on the source domain, significantly improving performance in deployment scenarios. • Validated the effectiveness of DAQuant through extensive experiments across challenging scenarios, including ultra-low-bit quantization and domain adaptation tasks, highlighting its robustness and applicability to edge devices.

Keywords:

Domain Aware Post-Training Quantization

Vision Transformers

Activation Smoothing

Learnable Activation Clipping

Domain Alignment

Domain aware post training quantization for vision transformers in deployment

Abstract

Journal

Organization