Fig. 3: Analysis of the time and memory consumption of the ABP and sum readouts on all datasets.
From: Modelling local and general quantum mechanical properties with attention-based pooling
Time spent training per epoch (minutes) and memory consumption (GB) of SchNet and DimeNet++ models with sum and ABP readouts (standard attention) for different datasets. Several configurations of ABP (hidden dimension per attention head, number of attention heads) are included for comparison, and all ABP readouts use 2 self attention blocks (SABs). Results are reported for 10 different runs of the same model configuration on a single NVIDIA V100 GPU with 32GB memory. QMugs models use a batch size of 64, while all other models use a batch size of 128. Error bars represent a 95% confidence interval.
