def _auto_step_lr_scheduler_per_batch(self, batch_idx: int, lr_scheduler: LRScheduler) -> None:
"""
This function aims at automatically step a LR scheduler. It should be called per batch.
"""
if lr_scheduler._step_mode == LRScheduler.StepMode.STEP_EVERY_BATCH:
lr_scheduler.step()
elif lr_scheduler._step_mode == LRScheduler.StepMode.STEP_EVERY_EPOCH:
mod = (batch_idx + 1) % len(self.training_loader)
if mod == 0 or mod < self.hvd_config.aggregation_frequency:
lr_scheduler.step()
def create_lr_scheduler(self, optimizer: torch.optim.Optimizer):
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=self.context.get_hparam("num_warmup_steps"),
num_training_steps=self.context.get_hparam("num_training_steps"),
)
return LRScheduler(scheduler, LRScheduler.StepMode.STEP_EVERY_BATCH)
def create_lr_scheduler(self, optimizer: torch.optim.Optimizer):
scheduler = torch.optim.lr_scheduler.StepLR(
optimizer,
self.context.get_hparam("reduce_every"),
gamma=self.context.get_hparam("lr_gamma"),
)
return LRScheduler(scheduler, LRScheduler.StepMode.STEP_EVERY_EPOCH)
def create_lr_scheduler(self, optimizer: torch.optim.Optimizer):
self.Lr2 = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, 100)
self.Lr1 = torch.optim.lr_scheduler.StepLR(optimizer, 1)
self.combined_lrs = MultiLRScheduler(self.Lr1, self.Lr2, optimizer)
# Because we are calling .step() ourselves in our MultiLRScheduler we need to
# set the StepMode to MANUAL_STEP.
return LRScheduler(self.combined_lrs, step_mode=LRScheduler.StepMode.MANUAL_STEP)
def create_lr_scheduler(self, optimizer: torch.optim.Optimizer):
"""
Required Method to use a learning rate scheduler
Returns: Determined scheduler object
"""
self.myLR = MyLR(optimizer, self.hparams)
step_mode = LRScheduler.StepMode.MANUAL_STEP
self._optimizer = self.context.get_optimizer()
return LRScheduler(self.myLR, step_mode=step_mode)
def _auto_step_lr_scheduler_per_batch(
self, batch_idx: int, lr_scheduler: pytorch.LRScheduler) -> None:
"""
This function aims at automatically step a LR scheduler. It should be called per batch.
"""
# Never step lr when we do not step optimizer.
if not self.context._should_communicate_and_update():
return
if lr_scheduler._step_mode == pytorch.LRScheduler.StepMode.STEP_EVERY_BATCH:
start_idx = batch_idx - self.context._aggregation_frequency + 1
for i in range(start_idx, batch_idx + 1):
if (i + 1) % lr_scheduler._frequency == 0:
lr_scheduler.step()
elif lr_scheduler._step_mode == pytorch.LRScheduler.StepMode.STEP_EVERY_OPTIMIZER_STEP:
if (batch_idx + 1) % lr_scheduler._frequency == 0:
lr_scheduler.step()
elif lr_scheduler._step_mode == pytorch.LRScheduler.StepMode.STEP_EVERY_EPOCH:
# We will step if the next optimizer step will land in the next epoch.
epoch_idx = self.get_epoch_idx(batch_idx)
next_steppable_batch = batch_idx + self.context._aggregation_frequency
next_batch_epoch_idx = self.get_epoch_idx(next_steppable_batch)
for e in range(epoch_idx, next_batch_epoch_idx):
if (e + 1) % lr_scheduler._frequency == 0:
lr_scheduler.step()
def create_lr_scheduler(self, optimizer: torch.optim.Optimizer):
"""
Required Method to use a learning rate scheduler
Returns: Determined scheduler object
Determined will handle the learning rate scheduler update based on the Determined
LRScheduler parameters. If step_every_batch or step_every_epoch is True, Determined will
handle the .step(). If both are false, the user will be in charge of calling .step().
"""
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=self.context.get_hparam("num_warmup_steps"),
num_training_steps=self.context.get_hparam("num_training_steps"),
)
return LRScheduler(scheduler, LRScheduler.StepMode.STEP_EVERY_BATCH)
def create_lr_scheduler(self, optimizer: torch.optim.Optimizer):
"""
Required Method to use a learning rate scheduler
Returns: Determined scheduler object
Determined will handle the learning rate scheduler update based on the Determined
LRScheduler parameters. If step_every_batch or step_every_epoch is True, Determined will
handle the .step(). If both are false, the user will be in charge of calling .step().
"""
self.myLR = MyLR(optimizer, self.context.get_hparams())
step_mode = LRScheduler.StepMode.MANUAL_STEP
if self.context.get_hparam("step_every_batch"):
step_mode = LRScheduler.StepMode.STEP_EVERY_BATCH
elif self.context.get_hparam("step_every_epoch"):
step_mode = LRScheduler.StepMode.STEP_EVERY_EPOCH
return LRScheduler(self.myLR, step_mode=step_mode)
def create_lr_scheduler(self, optimizer):
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=3,
gamma=0.1)
return LRScheduler(lr_scheduler,
step_mode=LRScheduler.StepMode.STEP_EVERY_EPOCH)
def create_lr_scheduler(self, optimizer):
self.scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer, self.context.get_hparam("cosine_annealing_epochs")
)
step_mode = LRScheduler.StepMode.MANUAL_STEP
return LRScheduler(self.scheduler, step_mode=step_mode)
def create_lr_scheduler(self, optimizer: torch.optim.Optimizer):
step_mode = LRScheduler.StepMode.STEP_EVERY_BATCH
scheduler = build_lr_scheduler(self.cfg, optimizer)
return LRScheduler(scheduler, step_mode=step_mode)
def create_lr_scheduler(self, optimizer):
self.scheduler = ModifyableLRSchedule(optimizer)
return LRScheduler(self.scheduler,
step_mode=LRScheduler.StepMode.MANUAL_STEP)
def create_lr_scheduler(self, optimizer):
self.scheduler = ModifyableLRSchedule(optimizer)
return LRScheduler(self.scheduler,
step_mode=LRScheduler.StepMode.STEP_EVERY_BATCH)
def create_lr_scheduler(self, optimizer):
self.scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer, self.context.get_hparam("train_epochs"))
step_mode = LRScheduler.StepMode.STEP_EVERY_EPOCH
return LRScheduler(self.scheduler, step_mode=step_mode)
