Table 1.

Details of GhostNet used. The kernel size is applied to all convolutions in the corresponding G-bneck. #mid represents the number of extended channels in the Ghost module. #out represents the number of output channels. SE stands for a plug-and-play attention module.

Input	Operator	Kernel Size	#Mid	#Out	SE	Stride
256 × 256 × 3	Conv2d	3 × 3	-	16	-	2
128 × 128 × 16	G-bneck	3 × 3	16	16	-	1
128 × 128 × 16	G-bneck	3 × 3	48	24	-	2
64 × 64 × 24	G-bneck	3 × 3	72	24	-	1
64 × 64 × 24	G-bneck	5 × 5	72	40	True	2
32 × 32 × 40	G-bneck	5 × 5	120	40	True	1
32 × 32 × 40	G-bneck	3 × 3	240	80	-	2
16 × 16 × 80	G-bneck	3 × 3	200	80	-	1
16 × 16 × 80	G-bneck	3 × 3	184	80	-	1
16 × 16 × 80	G-bneck	3 × 3	184	80	-	1
16 × 16 × 80	G-bneck	3 × 3	480	112	True	1
16 × 16 × 112	G-bneck	3 × 3	672	112	True	1
16 × 16 × 112	G-bneck	5 × 5	672	160	True	2
8 × 8 × 160	G-bneck	5 × 5	960	160	-	1
8 × 8 × 160	G-bneck	5 × 5	960	160	True	1
8 × 8 × 160	G-bneck	5 × 5	960	160	-	1
8 × 8 × 160	G-bneck	5 × 5	960	160	True	1