[Bug] accumulative_counts does not apply to parameter_schedulers

[HAOCHENYE]

The accurate cumulative pipeline should be as follows:

    loss = model(data)
    loss.backward()
    if idx % accumulative_counts == 0:
        optimizer.step()
        scheduler.step()
        optimizer.zero_grad()

However, the condition idx % accumulative_counts is only relevant to the OptimWrapper in this context:

mmengine/mmengine/optim/optimizer/optimizer_wrapper.py

Line 339 in a483dba

def should_update(self) -> bool:

Meanwhile, the schedulers continue to update the learning rate at their original frequency. In this case, the pipeline looks like this:

    loss = model(data)
    loss.backward()
    if idx % accumulative_counts == 0:
        optimizer.step()
        optimizer.zero_grad()
    scheduler.step()

We need to fix this.

[Y-T-G]

Has this been fixed?

[Y-T-G]

Does this bug apply if by_epoch=True for param_scheduler?

[HAOCHENYE]

Does this bug apply if by_epoch=True for param_scheduler?

Unfortunately, that's the case. We haven't identified a suitable solution yet, as tightly integrating the functions of OptimWrapper and scheduler isn't advisable.

[flytocc]

To modify ParamSchedulerHook?