ResearchTrend.AI
  • Papers
  • Communities
  • Events
  • Blog
  • Pricing
Papers
Communities
Social Events
Terms and Conditions
Pricing
Parameter LabParameter LabTwitterGitHubLinkedInBlueskyYoutube

© 2025 ResearchTrend.AI, All rights reserved.

  1. Home
  2. Papers
  3. 2411.15397
88
0

Efficient Online Inference of Vision Transformers by Training-Free Tokenization

23 November 2024
Leonidas Gee
Wing Yan Li
V. Sharmanska
Novi Quadrianto
    ViT
ArXivPDFHTML
Abstract

The cost of deploying vision transformers increasingly represents a barrier to wider industrial adoption. Existing compression requires additional end-to-end fine-tuning or incurs a significant drawback to runtime, thus making them ill-suited for online inference. We introduce the Visual Word Tokenizer\textbf{Visual Word Tokenizer}Visual Word Tokenizer (VWT), a training-free method for reducing energy costs while retaining performance and runtime. The VWT groups patches (visual subwords) that are frequently used into visual words while infrequent ones remain intact. To do so, intra-image or inter-image statistics are leveraged to identify similar visual concepts for compression. Experimentally, we demonstrate a reduction in wattage of up to 19% with only a 20% increase in runtime at most. Comparative approaches of 8-bit quantization and token merging achieve a lower or similar energy efficiency but exact a higher toll on runtime (up to 2×2\times2× or more). Our results indicate that VWTs are well-suited for efficient online inference with a marginal compromise on performance.

View on arXiv
Comments on this paper