# these were manually filled in.
subclass = getattr(cannot_be_inferred, name)
else:
# this points to the optimizer class in torch.optim (e.g. Adam)
torch_optimizer_class = getattr(torch.optim, name)
# dynamically create wrapper class, which has the same name as torch_class
subclass = type(
name,
# must subclass Optimizer to be added to the Registry
(
OptimizerConfig, ),
{},
)
# fill in optimizer parameters (except params)
make_config_class(torch_optimizer_class,
blacklist=["params"])(subclass)
subclass.__hash__ = param_hash
classes[name] = subclass
@OptimizerConfig.fill_union()
class Optimizer__Union(TaggedUnion):
@classmethod
def default(cls, **kwargs):
"""Return default factory for Optimizer (defaulting to Adam)."""
return (cls(Adam=classes["Adam"]()) if kwargs == {} else cls(
Adam=classes["Adam"](**kwargs)))
def make_optimizer_scheduler(self, params):
return self.value.make_optimizer_scheduler(params)
for module in cannot_be_inferred_modules:
if hasattr(module, name):
cannot_be_inferred_module = module
break
if cannot_be_inferred_module is not None:
# these were manually filled in.
subclass = getattr(cannot_be_inferred_module, name)
else:
torch_lr_scheduler_class = getattr(torch.optim.lr_scheduler, name)
subclass = type(
name,
# must subclass Optimizer to be added to the Registry
(
LearningRateSchedulerConfig, ),
{"__module__": __name__},
)
make_config_class(torch_lr_scheduler_class,
blacklist=["optimizer"])(subclass)
subclass.__hash__ = param_hash
classes[name] = subclass
@LearningRateSchedulerConfig.fill_union()
class LearningRateScheduler__Union(TaggedUnion):
def make_from_optimizer(
self, optimizer: torch.optim.Optimizer
) -> torch.optim.lr_scheduler._LRScheduler:
return self.value.make_from_optimizer(optimizer)