Table 2.

Results of compressing ResNet20, ResNet56, VGG16 on Cifar-10 and ResNet18 on Cifar-100. SOP represents the set of operators used in the reconstructed CNN. TOP1 refers to the TOP1 accuracy of the model on the test set. MA is the architecture of the compressed model, as shown in Figure 5, Figure 6 and Figure 7. Paras represents the number of parameters in the CNN, containing the number of parameters of all convolutional layers except the first convolutional layer and the last fully connected layer.

Model	Method	SOP	$\mathit{\lambda }$	PCR	FCR	Paras	TOP1 (%)	MA
ResNet20	He. [45]	-	-	0	0	0.27 M	91.25	-
	Ours	SOP2	0	2.87	2.56	0.11 M	91.6	C.1
		SOP1	$1.5\times {10}^{-3}$	4.91	4.19	0.06 M	90.35	C.2
		SOP2	$1.5\times {10}^{-3}$	6.48	5.61	0.047 M	90.15	C.3
ResNet56	He. [45]	-	-	0	0	0.85 M	93.03	-
	Li. [18]	-	-	1.16	1.38	-	93.06	-
	Dug. [52]	-	-	-	2.12	-	92.72	-
	Ours	SOP3	0	3.51	3.09	0.24 M	93.75	C.7
		SOP1	0	4.4	3.37	0.19 M	92.5	C.6
		SOP1	$1.5\times {10}^{-3}$	5.25	5.96	0.17 M	91.96	C.5
		SOP2	$1.5\times {10}^{-3}$	6.59	7.94	0.14 M	91.22	C.4
VGG16	Simon. [53]	-	-	0	0	16.3 M	93.25	-
	Li [18]	-	-	2.78	1.52	-	93.4	-
	Dug. [52]	-	-	17.12	3.15	-	92.85	-
	Ours	SOP5	0	2.82	3.71	5.79 M	94.65	C.8
		SOP4	0	3.85	2.61	4.23 M	93.95	C.9
		SOP6	$1.5\times {10}^{-3}$	15.1	15.6	1.08 M	92.35	C.10
ResNet18	He. [45]		-	0	0	11 M	75.05	-
	Ours	SOP7	0	2.39	2.23	4.61 M	74.5	C.11
		SOP7	$1.5\times {10}^{-3}$	2.44	2.31	4.5 M	74.2	C.12
		SOP8	0	5.27	2.97	2.08 M	74.85	C.13
		SOP8	$1.5\times {10}^{-3}$	4.66	3.98	2.36 M	73.6	C.14