Table 2.

Time and resource utilization of different attention types for ADN on Metr-LA.

Model	No. Parameters	Training Time (s/epoch)	Inference Time (ms/sample)	Peak GPU Usage (GB)
ADN-DA	331 K	30	7.8	9.9
ADN-GA	331 K	46	1.9	4.7
ADN-RA	324 K	50	8.1	11
ADN-FA	331 K	23	2.4	4.5
ADN-EA	331 K	27	1.7	5.5
ADN-LA	341 K	23	2.1	4.5
ADN-FV	330 K	25	2.2	4.7