Computational Intelligence and Neuroscience / 2022 / Article / Tab 2 / Research Article
Differentiable Network Pruning via Polarization of Probabilistic Channelwise Soft Masks Table 2 Comparison results on the CIFAR10 dataset with VGG16, ResNet32, ResNet56, and ResNet110. Acc
is the accuracy drop of the pruned model compared to the baseline model. FLOPs
represent the pruning rate of FLOPs.
Network Method Baseline acc (%) Pruned Acc FLOPs VGG-16 HRank [11 ] 93.96 92.34 1.62 65.30 SCP [14 ] 93.85 93.79 0.06 66.23 PPSM (ours) 93.72 93.78 −0.06 66.20 ResNet32 LFPC [46 ] 92.63 92.12 0.51 52.60 Wang et al. [47 ] 93.18 93.27 −0.09 49.00 PPSM (ours) 93.19 93.31 −0.12 53.27 LRF [27 ] 92.49 92.54 −0.05 62.00 MainDP [15 ] 92.66 92.15 0.51 63.20 PPSM (ours) 93.19 93.28 −0.09 64.35 ResNet56 Zhuang et al. [28 ] 93.80 93.83 −0.03 47.00 HRank [11 ] 93.26 93.17 0.09 50.00 LFPC [46 ] 93.59 93.34 0.25 52.90 DMC [13 ] 93.62 93.69 −0.07 50.00 SRR-GR [48 ] 93.38 93.75 −0.37 53.80 SCP [14 ] 93.69 93.23 0.46 51.50 DPFPS [49 ] 93.81 93.20 0.61 52.86 Wang et al. [47 ] 93.69 93.76 −0.07 50.00 PPSM (ours) 93.44 93.57 −0.13 54.60 HRank [11 ] 93.26 90.72 2.54 74.10 LRF-60 [27 ] 93.45 93.19 0.26 73.90 PPSM (ours) 93.44 93.22 0.22 75.62 ResNet110 HRank [11 ] 93.50 92.65 0.85 68.60 LFPC [46 ] 93.68 93.79 −0.11 60.30 LRF [27 ] 93.76 94.34 −0.58 62.60 PPSM (ours) 93.60 93.83 -0.23 68.70
The bold values are given to highlight the best-performing method in each performance metric.
