Differentiable Network Pruning via Polarization of Probabilistic Channelwise Soft Masks

Table 3

Comparison results of ResNet50 and MobileNet v2 on ImageNet. Pruned top-1 and pruned top-5 denote the top-1 and top-5 accuracy after the pruning. Top-1 and top-5 denote the accuracy drop of the pruned model when compared to the baseline model.

Network	Method	Pruned top-1(%)	Top-1 (%)	Pruned top-5(%)	Top-5 (%)	FLOPs (%)
ResNet50	GAL [16]	71.80	4.35	90.82	2.05	55.01
	Zhuang et al. [28]	75.63	0.52	—	—	54.00
	DMCP [18]	76.20	0.40	—	—	46.34
	DMC [13]	75.35	0.80	92.49	0.38	55.00
	HRank [11]	74.98	1.17	92.33	0.54	43.77
	GBN [17]	75.18	0.67	92.41	0.25	55.06
	SRR-GR [48]	75.11	1.02	92.35	0.51	55.10
	SCP [14]	75.27	0.62	92.30	0.68	54.30
	SCOP [50]	75.26	0.89	92.53	0.34	54.60
	LRF [27]	75.71	0.50	92.80	0.02	56.40
	DPFPS [49]	75.55	0.60	92.54	0.33	46.20
	PPSM (ours)	75.78	0.35	92.83	0.03	53.07
	GAL [16]	69.88	6.27	89.75	3.12	61.37
	HRank [11]	71.98	4.17	91.01	1.86	62.10
	CHIP [51]	75.26	0.89	92.53	0.34	62.80
	PPSM (ours)	75.43	0.70	92.54	0.32	65.91
	GAL [16]	69.31	6.84	89.12	3.75	72.86
	HRank [11]	69.10	7.05	89.58	3.29	76.04
	DMCP [18]	74.40	2.20	—	—	73.17
	CHIP [51]	73.30	2.85	91.48	1.39	76.70
	PPSM (ours)	74.59	1.54	92.30	0.56	72.43
MobileNet v2	AMC [7]	70.80	1.00	—	—	27.00
	Metapruning [52]	71.20	0.80	—	—	27.00
	DPFPS [49]	71.10	0.90	—	—	24.89
	PPSM (ours)	71.43	0.45	89.92	0.37	28.69

The bold values are given to highlight the best-performing method in each performance metric.