def on_stage_start(self, state: State) -> None:
"""Stage start hook.
Args:
state (State): current state
"""
self.reduced_metric = self.reduced_metric or state.main_metric
scheduler = state.get_attr(
key="scheduler", inner_key=self.scheduler_key
)
assert scheduler is not None
self._scheduler = scheduler
if self.mode is None:
if isinstance(scheduler, BatchScheduler):
self.mode = "batch"
else:
self.mode = "epoch"
if (
isinstance(scheduler, OneCycleLRWithWarmup)
and self.mode == "batch"
):
scheduler.reset()
assert self.mode is not None
def on_stage_start(self, state: State):
optimizer = state.get_attr(
key="optimizer", inner_key=self.optimizer_key
)
assert optimizer is not None
self._optimizer = optimizer
self.init_lr = optimizer.defaults["lr"]
def on_stage_start(self, state: State):
"""
Checks that the current stage has correct optimizer
"""
self._optimizer = state.get_attr(key="optimizer",
inner_key=self.optimizer_key)
assert self._optimizer is not None
def on_stage_start(self, state: State):
"""
Checks that the current stage has correct criterion
"""
criterion = state.get_attr(key="criterion",
inner_key=self.criterion_key)
assert criterion is not None
self._criterion = criterion
def on_stage_start(self, state: State) -> None:
"""Stage start hook.
Args:
state (State): current state
"""
optimizer = state.get_attr(key="optimizer",
inner_key=self.optimizer_key)
assert optimizer is not None
self._optimizer = optimizer
self.init_lr = optimizer.defaults["lr"]