update the lr by iteration in epochbasedrunner

[Jacobew]

Hi, how can I update the learning rate by iteration but not by epoch when using an epochbasedrunner?

My other question is: which part of the codes should I modify to substitute epochbasedrunner with iterbasedrunner?

Looking forward for the reply :)

[zachcoleman]

Set the by_epoch argument in the LR config to False for iter-based updates like:

lr_config = dict(
    policy="step",
    warmup="linear",
    warmup_iters=500,
    warmup_ratio=0.001,
    step=[70_000],
    by_epoch=False,
)

[zachcoleman]

In the config you can also specify the runner type like:

runner = dict(type="IterBasedRunner", max_iters=200_000)