Fig. 3

Datasets of loops and non-loops, and performance of DeepMILO on these datasets. Numbers inside brackets in c are AUPRCs. a Different types of non-loops for training and testing the loop model. b Details of DeepMILO; combining pretrained anchor model and pretrained anchor orientation model helped training to converge faster. c Performance of DeepMILO on different test sets. Proportions of positive samples are 0.43, 0.61, 0.72, 0.072, and 0.063 for non-loop types 1, 2, 3, 4, and 5 datasets, respectively