Question regarding the learning rate update strategy

[zhitao-guo]

Thank you for your beautiful work~

I noticed that the learning rate is updated after every batch by default, which seems different from the more common strategies (update lr every epoch).

mind-vis/code/config.py

Line 73 in 08d5386

self.accum_iter = 1

mind-vis/code/sc_mbm/trainer.py

Line 63 in 08d5386

if data_iter_step % accum_iter == 0:

Could you please explain the reason behind this approach? Additionally, if possible, could you provide some useful references or links on this topic? Looking forward to your reply~