Table 1.

Details of the EEGNet architecture.

Layer	Layer Type	Filters	Size	Parameters	Output Dimension	Activation	Mode
1	Input				(C, T)
	Reshape				(1, C, T)
	Conv2D	F	(1,64)	64*F	(F, C, T)	Linear	same
	BatchNorm			2*F	(F, 1, T)
	DepthwiseConv2D	F	(C,1)	C*F	(F, 1, T)	Linear	valid
	BatchNorm				(F, 1, T)
	Activation				(F, 1, T)	ELU
	SpatialDropout2D				(F, 1, T)
2	SeparableConv2D	F	(1,8)	$8\ast F+F^2$	(F, 1, T)	Linear	same
	BatchNorm			2*F	(F, 1, T)
	Activation				(F, 1, T)	ELU
	AveragePool2D		(1,4)		(F, 1, T//4)
	SpatialDropout2D				(F, 1, T//4)
3	SeparableConv2D	2*F	(1,8)	$2\ast F\ast 8+{\left(2\ast F\right)}^2$	(2*F, 1, T//4)	Linear	same
	BatchNorm			2*F	(2*F, 1, T//4)
	Activation				(2*F, 1, T//4)	ELU
	AveragePool2D		(1,4)		(2*F, 1, T//16)
	SpatialDropout2D				(2*F, 1, T//16)
4	Flatten				(2*F, 1, T//16)
	Dense				N	Softmax

C = number of channels, T = number of time points, F = number of filters and N = number of classes.