MBS: Macroblock Scaling for CNN Model Reduction

TL;DR

The paper introduces Macroblock Scaling (MBS), a simple and efficient algorithm for reducing CNN model sizes by adaptively pruning macroblocks based on information redundancy, achieving significant compression with minimal accuracy loss.

Contribution

MBS is a novel, scalable macroblock-level pruning method that outperforms existing optimization-based approaches in CNN model reduction.

Findings

Reduces model size of MobileNetV2 by 25.03%

Achieves 72.71% reduction in ResNet-1202

Maintains accuracy with negligible degradation

Abstract

In this paper we propose the macroblock scaling (MBS) algorithm, which can be applied to various CNN architectures to reduce their model size. MBS adaptively reduces each CNN macroblock depending on its information redundancy measured by our proposed effective flops. Empirical studies conducted with ImageNet and CIFAR-10 attest that MBS can reduce the model size of some already compact CNN models, e.g., MobileNetV2 (25.03% further reduction) and ShuffleNet (20.74%), and even ultra-deep ones such as ResNet-101 (51.67%) and ResNet-1202 (72.71%) with negligible accuracy degradation. MBS also performs better reduction at a much lower cost than the state-of-the-art optimization-based methods do. MBS's simplicity and efficiency, its flexibility to work with any CNN model, and its scalability to work with models of any depth make it an attractive choice for CNN model size reduction.