raw_config = self.config.to_dict()
raw_config.type = self.config.type
map_dict = LrSchedulerMappingDict()
self.map_config = ConfigBackendMapping(
map_dict.type_mapping_dict, map_dict.params_mapping_dict).backend_mapping(raw_config)
self._cls = ClassFactory.get_cls(ClassType.LR_SCHEDULER, self.map_config.type)
def __call__(self, optimizer=None, epochs=None, steps=None):
"""Call lr scheduler class."""
params = self.map_config.get("params", {})
logging.debug("Call LrScheduler. name={}, params={}".format(self._cls.__name__, params))
setattr(self._cls, "by_epoch", True)
if hasattr(self.config, "by_epoch"):
setattr(self._cls, "by_epoch", self.config.by_epoch)
try:
if params:
return self._cls(optimizer, **params)
else:
return self._cls(optimizer)
except Exception as ex:
logging.error("Failed to call LrScheduler name={}, params={}".format(self._cls.__name__, params))
raise ex
if zeus.is_torch_backend():
import torch.optim.lr_scheduler as torch_lr
ClassFactory.register_from_package(torch_lr, ClassType.LR_SCHEDULER)
try:
if params:
cls_obj = self._cls(**params) if isclass(
self._cls) else partial(self._cls, **params)
else:
cls_obj = self._cls() if isclass(self._cls) else partial(
self._cls)
if zeus.is_torch_backend() and TrainerConfig().cuda:
cls_obj = cls_obj.cuda()
return cls_obj
except Exception as ex:
logging.error("Failed to call Loss name={}, params={}".format(
self._cls.__name__, params))
raise ex
if zeus.is_torch_backend():
import torch.nn as torch_nn
import timm.loss as timm_loss
ClassFactory.register_from_package(torch_nn, ClassType.LOSS)
ClassFactory.register_from_package(timm_loss, ClassType.LOSS)
elif zeus.is_tf_backend():
import tensorflow.compat.v1.losses as tf_loss
ClassFactory.register_from_package(tf_loss, ClassType.LOSS)
elif zeus.is_ms_backend():
import mindspore.nn.loss as ms_loss
ClassFactory.register_from_package(ms_loss, ClassType.LOSS)
return pfms
def reset(self):
"""Reset states for new evaluation after each epoch."""
for val in self.mdict.values():
val.reset()
@property
def results(self):
"""Return metrics results."""
res = {}
for name, metric in self.mdict.items():
res.update(metric.result)
return res
@property
def objectives(self):
"""Return objectives results."""
return {name: self.mdict.get(name).objective for name in self.mdict}
def __getattr__(self, key):
"""Get a metric by key name.
:param key: metric name
:type key: str
"""
return self.mdict[key]
ClassFactory.register_from_package(metrics, ClassType.METRIC)
optimizer = hvd.DistributedOptimizer(optimizer,
named_parameters=model.named_parameters(),
compression=hvd.Compression.none)
elif zeus.is_tf_backend():
optimizer = dynamic_optimizer(self.optim_cls, **params)
if distributed:
optimizer = hvd.DistributedOptimizer(optimizer) if zeus.is_gpu_device() else \
NPUDistributedOptimizer(optimizer)
elif zeus.is_ms_backend():
learnable_params = [param for param in model.trainable_params() if param.requires_grad]
optimizer = self.optim_cls(learnable_params, **params)
return optimizer
except Exception as ex:
logging.error("Failed to call Optimizer name={}, params={}".format(self.optim_cls.__name__, params))
raise ex
if zeus.is_torch_backend():
import torch.optim as torch_opt
ClassFactory.register_from_package(torch_opt, ClassType.OPTIMIZER)
elif zeus.is_tf_backend():
import tensorflow.compat.v1.train as tf_train
ClassFactory.register_from_package(tf_train, ClassType.OPTIMIZER)
elif zeus.is_ms_backend():
import mindspore.nn.optim as ms_opt
ClassFactory.register_from_package(ms_opt, ClassType.OPTIMIZER)
