def lightning_scheduler_dict_to_det(lrs: dict) -> _LRScheduler:
"""
wrap user defined lr_scheduler and switch the attached optimizer with the
wrapped version.
input_dict = {
'scheduler': None,
'name': None, # no custom name
'interval': 'epoch', # after epoch is over
'frequency': 1, # every epoch/batch
'reduce_on_plateau': False, # most often not ReduceLROnPlateau scheduler
'monitor': monitor, # value to monitor for ReduceLROnPlateau
'strict': True, # enforce that the monitor exists for ReduceLROnPlateau
}
"""
if lrs["reduce_on_plateau"]:
raise InvalidModelException(
"LRScheduler reduce_on_plateau is not supported")
if lrs["monitor"] is not None:
raise InvalidModelException(
"LRScheduler monitor is not supported")
step_mode = (LRScheduler.StepMode.STEP_EVERY_EPOCH
if lrs["interval"] == "epoch" else
LRScheduler.StepMode.STEP_EVERY_BATCH)
opt_key = cast(Optimizer,
getattr(lrs["scheduler"], "optimizer", None))
wrapped_opt = optimizers_dict.get(opt_key)
if wrapped_opt is None:
raise InvalidModelException(
"An LRScheduler is returned in `configure_optimizers` without having "
"returned the optimizer itself. Please follow PyTorchLightning's documenation"
"to make sure you're returning one of the expected values."
"- Single optimizer.\n"
"- List or Tuple - List of optimizers.\n"
"- Two lists - The first list has multiple optimizers, the second a list of"
"LRSchedulers (or lr_dict).\n"
"- Dictionary, with an ‘optimizer’ key, and (optionally) a ‘lr_scheduler’ key"
"whose value is a single LR scheduler or lr_dict.\n"
"- Tuple of dictionaries as described, with an optional ‘frequency’ key.\n"
)
check.check_isinstance(
lrs["scheduler"].optimizer,
Optimizer,
"A returned LRScheduler from `configure_optimizers` is "
"missing the optimizer attribute.",
)
# switch the user's unwrapped optimizer with the wrapped version.
lrs["scheduler"].optimizer = wrapped_opt
return self._pls.context.wrap_lr_scheduler(
lrs["scheduler"], step_mode, frequency=lrs["frequency"])
def check_compatibility(lm: pl.LightningModule) -> None:
prefix = "Unsupported usage in PLAdapter: "
unsupported_members = {
"backward",
"get_progress_bar_dict",
"manual_backward",
"on_fit_end",
"on_fit_start",
"on_load_checkpoint",
"on_pretrain_routine_end",
"on_pretrain_routine_start",
"on_save_checkpoint",
"on_test_batch_end",
"on_test_batch_start",
"on_test_epoch_end",
"on_test_epoch_start",
"on_train_epoch_end",
"optimizer_step",
"optimizer_zero_grad",
"setup",
"tbptt_split_batch",
"teardown",
"test_dataloader",
"test_epoch_end",
"test_step",
"test_step_end",
"training_step_end",
"transfer_batch_to_device",
"validation_step_end",
}
members = inspect.getmembers(lm, predicate=inspect.ismethod)
overridden_members = set(
map(lambda m: m[0], filter(lambda m: is_overridden(m[0], lm), members))
)
matches = unsupported_members & overridden_members
if len(matches) > 0:
raise InvalidModelException(prefix + f"{matches}")
for member in overridden_members:
if has_param(getattr(lm, member), "dataloader_idx"):
raise InvalidModelException(
prefix
+ f'multiple dataloaders and `dataloader_idx` are not supported in "{member}"'
)
if has_param(lm.training_step, "hiddens", 4):
raise InvalidModelException(prefix + '`hiddens` argument in "training_step"')
if lm.trainer is not None:
raise InvalidModelException(prefix + "Lightning Trainer")
def lm_log_dict(a_dict: Dict, *args: Any, **kwargs: Any) -> None:
if len(args) != 0 or len(kwargs) != 0:
raise InvalidModelException(
f"unsupported arguments to LightningModule.log {args} {kwargs}"
)
for metric, value in a_dict.items():
if type(value) == int or type(value) == float:
writer.add_scalar(metric, value, context.current_train_batch())
def lightning_scheduler_dict_to_det(lrs: dict) -> _LRScheduler:
"""
input_dict = {
'scheduler': None,
'name': None, # no custom name
'interval': 'epoch', # after epoch is over
'frequency': 1, # every epoch/batch
'reduce_on_plateau': False, # most often not ReduceLROnPlateau scheduler
'monitor': monitor, # value to monitor for ReduceLROnPlateau
'strict': True, # enforce that the monitor exists for ReduceLROnPlateau
}
"""
if lrs["reduce_on_plateau"]:
raise InvalidModelException("LRScheduler reduce_on_plateaue is not supported")
if lrs["monitor"] is not None:
raise InvalidModelException("LRScheduler monitor is not supported")
step_mode = (
LRScheduler.StepMode.STEP_EVERY_EPOCH
if lrs["interval"] == "epoch"
else LRScheduler.StepMode.STEP_EVERY_BATCH
)
return context.wrap_lr_scheduler(lrs["scheduler"], step_mode)
def _build_train_args(self, batch: TorchData, batch_idx: int,
opt_idx: int) -> List[Any]:
# taken from pytorch_lightning
args = [batch, batch_idx]
if len(self._pls.optimizers) > 1:
if has_param(self._pls.lm.training_step, "optimizer_idx"):
args.append(opt_idx)
else:
num_opts = len(self._pls.optimizers)
raise InvalidModelException(
f"Your LightningModule defines {num_opts} optimizers but "
f'training_step is missing the "optimizer_idx" argument.')
return args