Table 1.

Autoencoder architecture summary.

Layer		Output shape	Parameter
Residual unit	Kernel 3, stride 3, padding 1	$4\times 341$	248
Residual unit	Kernel 3, stride 3, padding 1	$16\times 113$	3344
Residual unit	Kernel 3, stride 3, padding 1	$32\times 37$	13,216
Residual unit	Kernel 3, stride 3, padding 1	$64\times 12$	52,032
Flatten		$1\times 768$	–
Linear feed-forward		$1\times 16=dim\mathrm{\Xi }$	12,304
Linear feed-forward		$1\times 4608$	78,336
Reshape		$128\times 36$	–
Transposed conv. layer	Kernel 7, stride 3	$64\times 112$	57,664
Transposed conv. layer	Kernel 6, stride 3	$32\times 339$	12,448
Transposed conv. layer	Kernel 6, stride 3	$16\times 1020$	3152
Transposed conv. layer	Kernel 3, stride 1, padding 1	$8\times 1020$	424
Transposed conv. layer	Kernel 3, stride 1, padding 1	$4\times 1020$	116
Transposed conv. layer	Kernel 5, stride 1	$1\times 1024=dim\mathcal{U}$	29