How to use schedulers without epochs

[fsolgui]

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 for data in loader without any prior for epoch.

In this situation, how would you approach using schedulers, specifically the cosine scheduler?

In issue #1168 @rwightman commented:

BUT I've held off because I was thinking of making more significant changes to the schedulers themselves (and the epochs / steps) interface...

Are there any planned changes to schedulers to make it easier to work without epochs?

[rwightman]

@fsolgui the example train script does it when the --sched-on-updates flag 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 the train.py file...

Though the concept of epoch doesn't completely disappear if you use the scheduler factory, it translates epochs -> steps and requires updates_per_epoch to 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 set t_in_epochs=False that will do a step schedule as well. The scheduler.step_update() is called with current step count after each step instead of calling .step() per epoch