Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Comprehensive Study of Structural Pruning for Vision Models

Changhao Li, Haoling Li, Mengqi Xue, Gongfan Fang, Sheng Zhou, Zunlei Feng, Huiqiong Wang, Mingli Song, Jie Song

TL;DR

This work addresses the fragmented evaluation of structural pruning by introducing PruningBench, a standardized benchmark with a four-stage framework (sparsifying, grouping via DepGraph, pruning, finetuning) and reproducible leaderboards. It systematically evaluates 16 pruning methods across CNNs and ViTs on CIFAR100, ImageNet, and COCO, using both sparsifying-stage and pruning-stage tracks and a protected global pruning scheme to stabilize high-speedups. Key findings show that no single method dominates across all settings, with weight-norm approaches frequently performing well, while architecture and dataset choice significantly shape results; ViTs remain comparatively harder to prune without larger accuracy losses. The platform and interfaces promote fair comparisons and rapid integration of new methods, facilitating broader adoption and guiding future research toward more robust, generalizable pruning strategies.

Abstract

Structural pruning has emerged as a promising approach for producing more efficient models. Nevertheless, the community suffers from a lack of standardized benchmarks and metrics, leaving the progress in this area not fully comprehended. To fill this gap, we present the first comprehensive benchmark, termed PruningBench, for structural pruning. PruningBench showcases the following three characteristics: 1) PruningBench employs a unified and consistent framework for evaluating the effectiveness of diverse structural pruning techniques; 2) PruningBench systematically evaluates 16 existing pruning methods, encompassing a wide array of models (e.g., CNNs and ViTs) and tasks (e.g., classification and detection); 3) PruningBench provides easily implementable interfaces to facilitate the implementation of future pruning methods, and enables the subsequent researchers to incorporate their work into our leaderboards. We provide an online pruning platform for customizing pruning tasks and reproducing all results in this paper. Leaderboard results can also be available.

A Comprehensive Study of Structural Pruning for Vision Models

TL;DR

This work addresses the fragmented evaluation of structural pruning by introducing PruningBench, a standardized benchmark with a four-stage framework (sparsifying, grouping via DepGraph, pruning, finetuning) and reproducible leaderboards. It systematically evaluates 16 pruning methods across CNNs and ViTs on CIFAR100, ImageNet, and COCO, using both sparsifying-stage and pruning-stage tracks and a protected global pruning scheme to stabilize high-speedups. Key findings show that no single method dominates across all settings, with weight-norm approaches frequently performing well, while architecture and dataset choice significantly shape results; ViTs remain comparatively harder to prune without larger accuracy losses. The platform and interfaces promote fair comparisons and rapid integration of new methods, facilitating broader adoption and guiding future research toward more robust, generalizable pruning strategies.

Abstract

Structural pruning has emerged as a promising approach for producing more efficient models. Nevertheless, the community suffers from a lack of standardized benchmarks and metrics, leaving the progress in this area not fully comprehended. To fill this gap, we present the first comprehensive benchmark, termed PruningBench, for structural pruning. PruningBench showcases the following three characteristics: 1) PruningBench employs a unified and consistent framework for evaluating the effectiveness of diverse structural pruning techniques; 2) PruningBench systematically evaluates 16 existing pruning methods, encompassing a wide array of models (e.g., CNNs and ViTs) and tasks (e.g., classification and detection); 3) PruningBench provides easily implementable interfaces to facilitate the implementation of future pruning methods, and enables the subsequent researchers to incorporate their work into our leaderboards. We provide an online pruning platform for customizing pruning tasks and reproducing all results in this paper. Leaderboard results can also be available.
Paper Structure (19 sections, 3 figures, 5 tables)

This paper contains 19 sections, 3 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. PruningBench Framework
  4. PruningBench Settings
  5. Benchmark Results and Discussions
  6. Benchmark Results
  7. More Discussions
  8. Conclusion and Future Work
  9. Appendix
  10. Hyperparameters
  11. Pruning Step
  12. Hyperparameters of Pretraining
  13. Hyperparameters of Finetuning Stage
  14. Hyperparameters of Sparsifying Stage.
  15. Unified Interfaces
  16. ...and 4 more sections

Figures (3)

  • Figure 1: The framework of PruningBench, consisting of four steps: sparsifying, grouping, pruning and finetuning. Note that when benchmarking sparsifying regularizers (importance criteria), all other steps are fixed for fair comparisons.
  • Figure 2: The parameter curves of different models pruned by different importance criteria on CIFAR100 dataset. Details can be referred to the github leaderboards.
  • Figure 3: Results of MagnitudeL2 with different speedup ratios on ImageNet and CIFAR100.