• Main Navigation
  • Main Content
  • Sidebar

Russian Digital Libraries Journal

  • Home
  • About
    • About the Journal
    • Aims and Scopes
    • Themes
    • Editor-in-Chief
    • Editorial Team
    • Submissions
    • Open Access Statement
    • Privacy Statement
    • Contact
  • Current
  • Archives
  • Register
  • Login
  • Search
Published since 1998
ISSN 1562-5419
16+
Language
  • Русский
  • English

Search

Advanced filters

Search Results

Vit Quantization: CPU-Centric Analysis of the Trade-Off between Size and Speed

Amir Ramisovich Nigmatullin, Rustam Arifovich Lukmanov, Ahmad Taha
262-286
Abstract:

Using Vision Transformer (ViT) models in real medical practice – for example, in hospitals or diagnostic centers – is often difficult because doctors' work computers usually do not have powerful graphics processors (GPUs), and computing resources are limited. This work investigates a complete practical pipeline for model inference, aimed at reducing computational costs without significant loss of predictive performance.


The proposed approach combines several optimization techniques. First, knowledge distillation (KD) is used, where a compact student model learns to mimic the behavior of a larger, more accurate teacher model. Second, Exponential Moving Average (EMA) of the model weights is applied to stabilize training and improve generalization. Third, post-training INT8 quantization (PTQ) is explored to reduce model size and accelerate inference. Additionally, a simplified quantization-aware training variant (QAT-lite) is considered, where the effects of quantization are partially incorporated during fine-tuning.


Experiments are conducted on the ISIC dataset, which contains dermoscopic images of skin lesions. Model performance is evaluated using standard classification metrics, including accuracy, macro-averaged F1 score, and area under the ROC curve (ROC-AUC). CPU performance is also analyzed, including inference latency, throughput, memory consumption, and the final model size.


The results show that post-training INT8 quantization preserves performance close to the FP32 baseline while substantially reducing memory and computational requirements. In contrast, QAT-lite does not consistently provide reproducible improvements over PTQ.

Keywords: Vision Transformer, knowledge distillation, EMA, post-training quantization, quantization-aware training.
1 - 1 of 1 items
Information
  • For Readers
  • For Authors
  • For Librarians
Make a Submission
Current Issue
  • Atom logo
  • RSS2 logo
  • RSS1 logo

Russian Digital Libraries Journal

ISSN 1562-5419

Information

  • About the Journal
  • Aims and Scopes
  • Themes
  • Author Guidelines
  • Submissions
  • Privacy Statement
  • Contact
  • eLIBRARY.RU
  • dblp computer science bibliography

Send a manuscript

Authors need to register with the journal prior to submitting or, if already registered, can simply log in and begin the five-step process.

Make a Submission
About this Publishing System

© 2015-2026 Kazan Federal University; Institute of the Information Society