about train about sg_pr

[kxhit]

We randomly sample the positive pairs (e.g. <3m) and negative pairs (e.g. >20m) by a ratio (e.g. 1:100) for training. And for evaluation, we remove the easy positive pairs (e.g. <3m & timestamps < 30s). You can refer to the old version code or this repo.
Note that the ratio really matters when compared with other methods (Fig.6 in IROS'21 paper). Please keep all methods compared strictly on the same data distribution.