Blending Pruning Criteria for Convolutional Neural Networks

The advancement of convolutional neural networks (CNNs) on various vision applications has attracted lots attention. Yet the majority CNNs are unable to satisfy strict requirement for real-world deployment. To overcome this, recent popular network pruning is an effective method reduce redundancy models. However, ranking filters according their “importance” different criteria may be inconsistent. One filter could important a certain criterion, while it unnecessary another one, which indicates that each criterion only partial view comprehensive “importance”. From this motivation, we propose novel framework integrate existing by exploring diversity. proposed contains two stages: Criteria Clustering and Filters Importance Calibration. First, condense via layerwise clustering based rank score. Second, within cluster, calibration factor adjust significance selected blending candidates search optimal Evolutionary Algorithm. Quantitative results CIFAR-100 ImageNet benchmarks show our outperforms state-of-the-art baselines, regrading compact model performance after pruning.