Fix a bug in learning rate scheduler for triplet loss

[Cysu]

Now it achieves slightly better result than Hermans, et al., reported under the "original" setting in Table 4.