Comb, Prune, Distill: Towards Unified Pruning for Vision Model Compression

TL;DR

This paper introduces a unified pruning framework called CPD that enables architecture- and task-agnostic model compression for vision models, combining hierarchical pruning, importance scoring, and knowledge distillation.

Contribution

The proposed CPD framework is the first to unify pruning across different architectures and tasks, addressing dependency issues and incorporating distillation for better retention.

Findings

Achieves up to 4.3x speedup in image classification with 1.8% accuracy loss.

Attains up to 1.89x speedup in segmentation with 5.1% mIoU loss.

Demonstrates generalizability across CNN and transformer models.

Abstract

Lightweight and effective models are essential for devices with limited resources, such as intelligent vehicles. Structured pruning offers a promising approach to model compression and efficiency enhancement. However, existing methods often tie pruning techniques to specific model architectures or vision tasks. To address this limitation, we propose a novel unified pruning framework Comb, Prune, Distill (CPD), which addresses both model-agnostic and task-agnostic concerns simultaneously. Our framework employs a combing step to resolve hierarchical layer-wise dependency issues, enabling architecture independence. Additionally, the pruning pipeline adaptively remove parameters based on the importance scoring metrics regardless of vision tasks. To support the model in retaining its learned information, we introduce knowledge distillation during the pruning step. Extensive experiments demonstrate the generalizability of our framework, encompassing both convolutional neural network (CNN) and transformer models, as well as image classification and segmentation tasks. In image classification we achieve a speedup of up to x4.3 with a accuracy loss of 1.8% and in semantic segmentation up to x1.89 with a 5.1% loss in mIoU.