Tensorflow+yolov3 training sample, the loss value is no longer convergent?

problem description

Hello everyone, I use tensorflow+yolov3 for target recognition, batch_size=32,random=0 (no multi-angle training), 5000 rounds of training, the loss value is about 4.1.
on this basis, I set random=1, which is also in the case of 5000 rounds, but the loss reaches 10. I continued to train on the basis of the original model and found that the loss converged to more than 4 o"clock midway, and then diverged globally and would not converge any more.

my inference

because batch_size is always convergent when set to 10, if set to 32, it may exceed the local best point.

my question

I would like to ask you heroes, in fact, is not trapped in the local optimal, in the local best caused by loss shock, I would like to ask what to do next?

MySQL Query : SELECT * FROM `codeshelper`.`v9_news` WHERE status=99 AND catid='6' ORDER BY rand() LIMIT 5
MySQL Error : Disk full (/tmp/#sql-temptable-64f5-7ae597-7e21.MAI); waiting for someone to free some space... (errno: 28 "No space left on device")
MySQL Errno : 1021
Message : Disk full (/tmp/#sql-temptable-64f5-7ae597-7e21.MAI); waiting for someone to free some space... (errno: 28 "No space left on device")
Need Help?