Neural Epitome Search for Architecture-Agnostic Network Compression

TL;DR

This paper introduces an auto-sampling method for CNN compression that learns sampling rules end-to-end, outperforming previous fixed-strategy methods like WSNet and achieving better accuracy and efficiency on 1D and 2D CNNs.

Contribution

The proposed auto-sampling approach automatically learns sampling strategies for network compression, enhancing performance over handcrafted methods and integrating differentiable rule learning.

Findings

Outperforms WSNet by 6.5% at the same compression ratio on 1D CNNs.

Achieves 1.47% higher accuracy than MobileNetV2 with 25% FLOPs reduction on ImageNet.

Surpasses some NAS-based methods like AMC and MnasNet in accuracy at similar compression levels.

Abstract

The recent WSNet [1] is a new model compression method through sampling filterweights from a compact set and has demonstrated to be effective for 1D convolutionneural networks (CNNs). However, the weights sampling strategy of WSNet ishandcrafted and fixed which may severely limit the expression ability of the resultedCNNs and weaken its compression ability. In this work, we present a novel auto-sampling method that is applicable to both 1D and 2D CNNs with significantperformance improvement over WSNet. Specifically, our proposed auto-samplingmethod learns the sampling rules end-to-end instead of being independent of thenetwork architecture design. With such differentiable weight sampling rule learning,the sampling stride and channel selection from the compact set are optimized toachieve better trade-off between model compression rate and performance. Wedemonstrate that at the same compression ratio, our method outperforms WSNetby6.5% on 1D convolution. Moreover, on ImageNet, our method outperformsMobileNetV2 full model by1.47%in classification accuracy with25%FLOPsreduction. With the same backbone architecture as baseline models, our methodeven outperforms some neural architecture search (NAS) based methods such asAMC [2] and MNasNet [3].