Attention-Based DSC-ConvLSTM for Multiclass Motor Imagery Classification

Table 1

Network structure.

Blocks	Layer	Filters	Output	Options
1	Input		[C, T]
	Reshape		[1, C, T]
	Conv_time	F1	[F1, C, T]	kernel_size = (1, 64), stride = (1, 1), padding = (0, F12)
	BatchNorm		[F1, C, T]	num_features = F1, eps = 0.001, momentum = 0.01
	Conv_spat	DF1	[DF1, 1, T]	kernel_size = (C, 1), stride=(1, 1),groups = F1
	BatchNorm		[DF1, 1, T]	num_features = DF1, eps = 0.001, momentum = 0.01
	Activation		[DF1, 1, T]	ELU
	MaxPool2D		[DF1, 1, T//3]	kernel_size = (1, 3),stride = (1, 3)
	Dropout		[DF1, 1, T//3]
2	DepthwiseConv2D	F2	[F2, 1, T//3]	kernel_size = (1, 16), stride = (1, 1), groups = F2
	PointwiseConv2D	DF2	[DF2, 1, T//3]	kernel_size = (1, 1), stride = (1, 1)
	BatchNorm		[DF2, 1, T//3]	num_features = DF2, eps = 0.001, momentum = 0.01
	Activation		[DF2, 1, T//3]	ELU
	MaxPool2D		[DF2, 1, T//9]	kernel_size = (1, 3), stride = (1, 3)
	Dropout
3	DepthwiseConv2D	F3	[F3, 1, T//9]	kernel_size = (1, 16), stride = (1, 1), groups = F3
	PointwiseConv2D	DF3	[DF3, 1, T//9]	kernel_size = (1, 1), stride = (1, 1)
	BatchNorm		[DF3, 1, T//9]	num_features = DF3, eps = 0.001, momentum = 0.01
	Activate		[DF3, 1, T//9]	ELU
	MaxPool2D		[DF3, 1, T//27]	kernel_size = (1, 3), stride=(1, 3)
	Dropout		[DF3, 1, T//27]
4	SeparableConv2D	F4	[F4, 1, T//27]	kernel_size = (1, 16), stride = (1, 1), groups = F4
	PointwiseConv2D	DF4	[DF4, 1, T//27]	kernel_size = (1, 1), stride = (1, 1)
	BatchNorm		[DF4, 1, T//27]	num_features = DF4, eps = 0.001, momentum = 0.01
	Activation		[DF4 ,1, T//27]	ELU
	MeanPool2D		[DF4, 1, T//81]	kernel_size = (1, 3), stride = (1, 3)
	Dropout		[DF4, 1, T//81]