-
|
I am currently working on generating synthetic data to train different models. The data is generated randomly each batch and therefore, the concept of epoch disappears. I have a single In this situation, how would you approach using schedulers, specifically the cosine scheduler? In issue #1168 @rwightman commented:
Are there any planned changes to schedulers to make it easier to work without epochs? |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment
-
|
@fsolgui the example train script does it when the Though the concept of epoch doesn't completely disappear if you use the scheduler factory, it translates epochs -> steps and requires |
Beta Was this translation helpful? Give feedback.
@fsolgui the example train script does it when the
--sched-on-updatesflag is set, you need to pass the # of steps per epoch to the schedule on creation and call a different step, can see use in thetrain.pyfile...Though the concept of epoch doesn't completely disappear if you use the scheduler factory, it translates epochs -> steps and requires
updates_per_epochto be set... though can use that if you set # epochs = 1 and updates_per_epoch the number of steps you need. If you use the schedulers directy and sett_in_epochs=Falsethat will do a step schedule as well. Thescheduler.step_update()is called with current step count after each step instead of calling.step()per epoch