Learning Scalable Model Soup on a Single GPU: An Efficient Subspace Training Strategy

TL;DR

This paper introduces MEHL-Soup, a memory-efficient and faster method for model soups that enables scalable hyperparameter tuning on a single GPU by formulating the problem as hyperplane optimization.

Contribution

The paper proposes MEHL-Soup, a novel hyperplane optimization approach for model soups that significantly reduces memory and computation costs, enabling single-GPU training.

Findings

MEHL-Soup outperforms Learned-Soup in test accuracy.

Memory usage is reduced by more than 13 times.

Soup construction speed is increased by 9 times.

Abstract

Pre-training followed by fine-tuning is widely adopted among practitioners. The performance can be improved by "model soups"~\cite{wortsman2022model} via exploring various hyperparameter configurations.The Learned-Soup, a variant of model soups, significantly improves the performance but suffers from substantial memory and time costs due to the requirements of (i) having to load all fine-tuned models simultaneously, and (ii) a large computational graph encompassing all fine-tuned models. In this paper, we propose Memory Efficient Hyperplane Learned Soup (MEHL-Soup) to tackle this issue by formulating the learned soup as a hyperplane optimization problem and introducing block coordinate gradient descent to learn the mixing coefficients. At each iteration, MEHL-Soup only needs to load a few fine-tuned models and build a computational graph with one combined model. We further extend MEHL-Soup to MEHL-Soup+ in a layer-wise manner. Experimental results on various ViT models and data sets show that MEHL-Soup(+) outperforms Learned-Soup(+) in terms of test accuracy, and also reduces memory usage by more than $13\times$. Moreover, MEHL-Soup(+) can be run on a single GPU and achieves $9\times$ speed up in soup construction compared with the Learned-Soup. The code is released at https://github.com/nblt/MEHL-Soup.