Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Merge-Friendly Post-Training Quantization for Multi-Target Domain Adaptation

Juncheol Shin, Minsang Seok, Seonggon Kim, Eunhyeok Park

TL;DR

This work tackles the challenge of combining quantization with training-free multi-target domain adaptation via model merging. It introduces HDRQ, a post-training quantization method that leverages noise-based Hessian regularization and weight distance regularization, along with noise-sampling rounding, to flatten the loss landscape and align merged models. A theoretical analysis of the error barrier extended to quantized merging motivates the regularizers and rounding strategy, providing a principled basis for HDRQ. Experiments on semantic segmentation and Office-Home MTDA demonstrate that HDRQ preserves single-model accuracy while substantially improving merging performance, especially at low bit-widths, enabling robust, real-time adaptive AI on devices with limited resources.

Abstract

Model merging has emerged as a powerful technique for combining task-specific weights, achieving superior performance in multi-target domain adaptation. However, when applied to practical scenarios, such as quantized models, new challenges arise. In practical scenarios, quantization is often applied to target-specific data, but this process restricts the domain of interest and introduces discretization effects, making model merging highly non-trivial. In this study, we analyze the impact of quantization on model merging through the lens of error barriers. Leveraging these insights, we propose a novel post-training quantization, HDRQ - Hessian and distant regularizing quantization - that is designed to consider model merging for multi-target domain adaptation. Our approach ensures that the quantization process incurs minimal deviation from the source pre-trained model while flattening the loss surface to facilitate smooth model merging. To our knowledge, this is the first study on this challenge, and extensive experiments confirm its effectiveness.

Merge-Friendly Post-Training Quantization for Multi-Target Domain Adaptation

TL;DR

This work tackles the challenge of combining quantization with training-free multi-target domain adaptation via model merging. It introduces HDRQ, a post-training quantization method that leverages noise-based Hessian regularization and weight distance regularization, along with noise-sampling rounding, to flatten the loss landscape and align merged models. A theoretical analysis of the error barrier extended to quantized merging motivates the regularizers and rounding strategy, providing a principled basis for HDRQ. Experiments on semantic segmentation and Office-Home MTDA demonstrate that HDRQ preserves single-model accuracy while substantially improving merging performance, especially at low bit-widths, enabling robust, real-time adaptive AI on devices with limited resources.

Abstract

Model merging has emerged as a powerful technique for combining task-specific weights, achieving superior performance in multi-target domain adaptation. However, when applied to practical scenarios, such as quantized models, new challenges arise. In practical scenarios, quantization is often applied to target-specific data, but this process restricts the domain of interest and introduces discretization effects, making model merging highly non-trivial. In this study, we analyze the impact of quantization on model merging through the lens of error barriers. Leveraging these insights, we propose a novel post-training quantization, HDRQ - Hessian and distant regularizing quantization - that is designed to consider model merging for multi-target domain adaptation. Our approach ensures that the quantization process incurs minimal deviation from the source pre-trained model while flattening the loss surface to facilitate smooth model merging. To our knowledge, this is the first study on this challenge, and extensive experiments confirm its effectiveness.

Paper Structure

This paper contains 17 sections, 12 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Quantization
  4. Domain Adaptation
  5. Analysis
  6. General error barrier case
  7. Regularization for Merge-Friendly Quantization
  8. Domain Adaptation case
  9. HDRQ: Hessian and Distance Regularizing Quantization
  10. Noise-based hessian regularization
  11. Weight distance regularization
  12. Handling ambiguity in rounding policy
  13. Experimental Results
  14. Semantic Segmentation
  15. Office Home
  16. ...and 2 more sections

Figures (3)

  • Figure 1: We propose a quantization scheme designed with future merging in mind. Our method ensures that networks are quantized to a more merge-friendly state, reducing the degradation induced by merging.
  • Figure 2: Visualization of loss surfaces quantized with each method is shown. ResNet-50 adapted from Real domain to Clipart domain (R $\xrightarrow{}$ C) is quantized to W4A8. HDRQ effectively regularize hessian with noise-based quantization, leading weights to flatter surface.
  • Figure 3: The distribution of harmonic mean accuracy for merging W4A8 quantized C$\xrightarrow[]{}$ R and C$\xrightarrow[]{}$ A models on the Office-Home dataset is presented. Our simple yet effective cosine similarity-based method, denoted as Advanced, successfully filters out low-quality weights, stabilizing merging outcomes.