Since the optimizer state is loaded from the checkpoint on resumption, changing any optimize values needs to occur after the checkpoint is loaded but before Trainer.fit is called.
This also might be doable with a new callback which fires on the after_load event, assuming all the callbacks are not replaced from the checkpoint callbacks. But it looks like that might happen per the docs.
Since the optimizer state is loaded from the checkpoint on resumption, changing any optimize values needs to occur after the checkpoint is loaded but before
Trainer.fitis called.This also might be doable with a new callback which fires on the
after_loadevent, assuming all the callbacks are not replaced from the checkpoint callbacks. But it looks like that might happen per the docs.