def __init__(self, parameters, mean=True, degree=None):
super(DistributedGradReducer, self).__init__(auto_prefix=False)
self.map_ = C.Map()
if degree is None:
self.degree = get_group_size()
else:
if not isinstance(degree, int) or degree <= 0:
raise ValueError(
"Parameter 'degree' in DistributedGradReducer should large than 0 and be int"
)
self.degree = degree
self.mean = mean
self.allreduce_filter = tuple(x.layerwise_parallel is False
for x in parameters)
is_parallel_optimizer = context.get_auto_parallel_context(
"enable_parallel_optimizer")
split_indices = auto_parallel_context(
).get_all_reduce_fusion_split_indices()
if is_parallel_optimizer and split_indices:
self.split_fusion = True
self.op_list = _init_allreduce_operators(len(parameters),
split_indices)
else:
self.split_fusion = False
self.allreduce = AllReduce().add_prim_attr('fusion', 1)
self.allgather = AllGather(GlobalComm.WORLD_COMM_GROUP)
ps_filter = lambda x: x.is_param_ps
self.ps_parameters = tuple(ps_filter(x) for x in parameters)
self.enable_parameter_server = any(self.ps_parameters)
def __init__(self,
params,
learning_rate=1e-3,
beta1=0.9,
beta2=0.999,
eps=1e-6,
weight_decay=0.0,
decay_filter=lambda x: 'beta' not in x.name and 'gamma' not in
x.name):
super(AdamWeightDecaySparse, self).__init__(learning_rate, params)
if self.is_group:
raise RuntimeError(
f"The {self.cls_name} optimizer cannot support group setting.")
_check_param_value(beta1, beta2, eps, weight_decay, self.cls_name)
self.beta1 = Tensor(np.array([beta1]).astype(np.float32))
self.beta2 = Tensor(np.array([beta2]).astype(np.float32))
self.eps = Tensor(np.array([eps]).astype(np.float32))
self.weight_decay_tensor = Tensor(
np.array([weight_decay]).astype(np.float32))
self.params = self.parameters
self.moments1 = self.params.clone(prefix="adam_m", init='zeros')
self.moments2 = self.params.clone(prefix="adam_v", init='zeros')
self.decay_flag = tuple(decay_filter(x) for x in self.params)
self.map = C.Map()
def __init__(self, parameters, mean=True, degree=None):
super(DistributedGradReducer, self).__init__(auto_prefix=False)
self.map_ = C.Map()
self.mul = P.Mul()
if degree is None:
self.degree = get_group_size()
else:
if not isinstance(degree, int) or degree <= 0:
raise ValueError("Parameter 'degree' in DistributedGradReducer should large than 0 and be int")
self.degree = degree
self.mean = mean
self.allreduce_filter = tuple(x.layerwise_parallel is False for x in parameters)
_init_optimizer_communication()
def __init__(self, parameters, mean=True, degree=None):
super(DistributedGradReducer, self).__init__(auto_prefix=False)
self.map_ = C.Map()
if degree is None:
self.degree = get_group_size()
else:
if not isinstance(degree, int) or degree <= 0:
raise ValueError(
"Parameter 'degree' in DistributedGradReducer should large than 0 and be int"
)
self.degree = degree
self.mean = mean
self.allreduce_filter = tuple(x.layerwise_parallel is False
for x in parameters)
self.opt_list = _init_allreduce_operators(len(parameters))
self.allgather = AllGather(GlobalComm.WORLD_COMM_GROUP)
ps_filter = lambda x: x.is_param_ps
self.ps_parameters = tuple(ps_filter(x) for x in parameters)
def __init__(self,
learning_rate,
parameters,
weight_decay=0.0,
loss_scale=1.0):
super(Optimizer, self).__init__(auto_prefix=False)
if parameters is not None and not isinstance(parameters, list):
parameters = list(parameters)
if not parameters:
raise ValueError("Optimizer got an empty parameter list.")
if not isinstance(parameters[0], (dict, Parameter)):
raise TypeError(
"Only a list of Parameter or dict can be supported.")
if isinstance(loss_scale, int):
loss_scale = float(loss_scale)
validator.check_value_type("loss_scale", loss_scale, [float],
self.cls_name)
validator.check_positive_float(loss_scale, "loss_scale", self.cls_name)
self.loss_scale = loss_scale
weight_decay = self._preprocess_weight_decay(weight_decay)
self._unique = True
self._target = context.get_context("device_target")
self.dynamic_lr = False
self.assignadd = None
self.global_step = None
self.is_group = False
self.is_group_lr = False
self.is_group_params_ordered = False
learning_rate = self._preprocess_single_lr(learning_rate)
if isinstance(parameters[0], dict):
self.is_group = True
self.group_params = []
self.group_lr = []
self.group_weight_decay = []
self._init_group_params(parameters, learning_rate, weight_decay)
# The final value of dynamic_lr can be determined after the process of parse_single_lr and init_group_params
if self.dynamic_lr:
self.assignadd = P.AssignAdd()
self.global_step = Parameter(initializer(0, [1], mindspore.int32),
name='global_step')
if self.is_group_lr:
if self.dynamic_lr:
self.learning_rate = CellList(self.group_lr)
else:
self.learning_rate = ParameterTuple(self.group_lr)
else:
self.learning_rate = self._build_single_lr(learning_rate,
'learning_rate')
if self.is_group:
self.parameters = ParameterTuple(self.group_params)
self.weight_decay = tuple(self.group_weight_decay)
self.weight_decay_tensor_tuple = tuple(
Tensor(x, mstype.float32) for x in self.group_weight_decay)
decay_filter = lambda x: x > 0
self.decay_flags = tuple(
decay_filter(x) for x in self.weight_decay)
self.exec_weight_decay = any(self.decay_flags)
else:
self.parameters = ParameterTuple(parameters)
self.weight_decay = weight_decay * loss_scale
self.weight_decay_tensor = Tensor(self.weight_decay,
mstype.float32)
decay_filter = lambda x: 'beta' not in x.name and 'gamma' not in x.name
self.decay_flags = tuple(decay_filter(x) for x in self.parameters)
self.exec_weight_decay = self.weight_decay > 0
# when a parameter has been unique, there is no need do another unique in optimizer.
for param in self.parameters:
if param.unique:
self._unique = False
break
ps_filter = lambda x: x.is_param_ps
self.ps_parameters = tuple(ps_filter(x) for x in self.parameters)
ps_cache_filter = lambda x: x.cache_enable
self.cache_enable = tuple(ps_cache_filter(x) for x in self.parameters)
self.reciprocal_scale = Tensor(1.0 / loss_scale, mstype.float32)
self.need_scale = loss_scale != 1.0
self.global_step_increase_tensor = Tensor(1, mstype.int32)
self.param_length = len(self.parameters)
self.map_ = C.Map()
if context.get_auto_parallel_context("enable_parallel_optimizer"):
if _get_parallel_mode(
) == ParallelMode.DATA_PARALLEL and context.get_context(
"device_target") == "Ascend":
self.use_parallel = True
elif _get_parallel_mode() == ParallelMode.DATA_PARALLEL \
and context.get_context("device_target") != "Ascend":
raise RuntimeError(
"Parallel optimizer only supports Ascend in data parallel mode."
)
elif _get_parallel_mode() in (ParallelMode.STAND_ALONE,
ParallelMode.HYBRID_PARALLEL):
raise RuntimeError(
"Parallel optimizer is not supported in {}.".format(
_get_parallel_mode()))
else:
self.use_parallel = False
else:
self.use_parallel = False
if self.use_parallel:
if self.cls_name not in ["Lamb", "AdamWeightDecay"]:
raise RuntimeError(
"Parallel optimizer does not support optimizer {}".format(
self.cls_name))
self.dev_num = _get_device_num()
if self.dev_num > self.param_length:
raise RuntimeError(
"Parallel optimizer can not be applied when the number of parameters {} is"
" less than the number of devices {}".format(
self.param_length, self.dev_num))
self.param_rank = self._get_parameter_group_id()
self.optim_filter = tuple(
map(lambda x: x == _get_global_rank(), self.param_rank))
self.param_names = []
for param in self.parameters:
self.param_names.append(param.name)
else:
self.optim_filter = (True, ) * self.param_length
def __init__(self, *args, **kwargs):
super(KungFuMomentum, self).__init__(*args, **kwargs)
self.map_ = C.Map()
self.all_reduce = KungFuAllReduce()
def __init__(self,
learning_rate,
parameters,
weight_decay=0.0,
loss_scale=1.0):
super(Optimizer, self).__init__(auto_prefix=False)
if parameters is not None and not isinstance(parameters, list):
parameters = list(parameters)
if not parameters:
raise ValueError("Optimizer got an empty parameter list.")
if not isinstance(parameters[0], (dict, Parameter)):
raise TypeError(
"Only a list of Parameter or dict can be supported.")
if isinstance(loss_scale, int):
loss_scale = float(loss_scale)
validator.check_value_type("loss_scale", loss_scale, [float],
self.cls_name)
validator.check_positive_float(loss_scale, "loss_scale", self.cls_name)
self.loss_scale = loss_scale
weight_decay = self._preprocess_weight_decay(weight_decay)
self.grad_centralization = False
self._unique = True
self._target = context.get_context("device_target")
self.dynamic_lr = False
self.assignadd = None
self.global_step = None
self.is_group = False
self.is_group_lr = False
self.is_group_params_ordered = False
learning_rate = self._preprocess_single_lr(learning_rate)
if isinstance(parameters[0], dict):
self.is_group = True
self.group_params = []
self.group_lr = []
self.group_weight_decay = []
self.group_grad_centralization = []
self._init_group_params(parameters, learning_rate, weight_decay,
self.grad_centralization)
# The final value of dynamic_lr can be determined after the process of parse_single_lr and init_group_params
if self.dynamic_lr:
self.assignadd = P.AssignAdd()
self.global_step = Parameter(initializer(0, [1], mindspore.int32),
name='global_step')
if self.is_group_lr:
self.learning_rate = CellList(
self.group_lr) if self.dynamic_lr else ParameterTuple(
self.group_lr)
else:
self.learning_rate = self._build_single_lr(learning_rate,
'learning_rate')
if self.is_group:
self.parameters = ParameterTuple(self.group_params)
self.weight_decay = tuple(self.group_weight_decay)
self.weight_decay_tensor_tuple = tuple(
Tensor(x, mstype.float32) for x in self.group_weight_decay)
decay_filter = lambda x: x > 0
self.decay_flags = tuple(
decay_filter(x) for x in self.weight_decay)
self.exec_weight_decay = any(self.decay_flags)
self.grad_centralization_flags = tuple(
self.group_grad_centralization)
else:
self.parameters = ParameterTuple(parameters)
self.weight_decay = weight_decay * loss_scale
self.weight_decay_tensor = Tensor(self.weight_decay,
mstype.float32)
decay_filter = lambda x: 'beta' not in x.name and 'gamma' not in x.name
self.decay_flags = tuple(decay_filter(x) for x in self.parameters)
self.exec_weight_decay = self.weight_decay > 0
# when a parameter has been unique, there is no need do another unique in optimizer.
for param in self.parameters:
if param.unique:
self._unique = False
break
ps_filter = lambda x: x.is_param_ps
self.ps_parameters = tuple(ps_filter(x) for x in self.parameters)
cache_filter = lambda x: x.cache_enable
self.cache_enable = tuple(cache_filter(x) for x in self.parameters)
self.reciprocal_scale = Tensor(1.0 / loss_scale, mstype.float32)
self.need_scale = loss_scale != 1.0
self.global_step_increase_tensor = Tensor(1, mstype.int32)
self.param_length = len(self.parameters)
self.map_ = C.Map()
self._use_parallel_optimizer()
def __init__(self, learning_rate, parameters, weight_decay=0.0, loss_scale=1.0):
super(Optimizer, self).__init__(auto_prefix=False)
if parameters is not None and not isinstance(parameters, list):
parameters = list(parameters)
if not parameters:
raise ValueError("Optimizer got an empty parameter list.")
if not isinstance(parameters[0], (dict, Parameter)):
raise TypeError("Only a list of Parameter or dict can be supported.")
if isinstance(loss_scale, int):
loss_scale = float(loss_scale)
validator.check_value_type("loss_scale", loss_scale, [float], self.cls_name)
validator.check_number_range("loss_scale", loss_scale, 0.0, float("inf"), Rel.INC_NEITHER, self.cls_name)
self.loss_scale = loss_scale
weight_decay = self._preprocess_weight_decay(weight_decay)
self.dynamic_lr = False
self.assignadd = None
self.global_step = None
self.is_group = False
self.is_group_lr = False
self.is_group_params_ordered = False
learning_rate = self._preprocess_single_lr(learning_rate)
if isinstance(parameters[0], dict):
self.is_group = True
self.group_params = []
self.group_lr = []
self.group_weight_decay = []
self._init_group_params(parameters, learning_rate, weight_decay)
# The final value of dynamic_lr can be determined after the process of parse_single_lr and init_group_params
if self.dynamic_lr:
self.assignadd = P.AssignAdd()
self.global_step = Parameter(initializer(0, [1], mindspore.int32), name='global_step')
if self.is_group_lr:
if self.dynamic_lr:
self.learning_rate = CellList(self.group_lr)
else:
self.learning_rate = ParameterTuple(self.group_lr)
else:
self.learning_rate = self._build_single_lr(learning_rate, 'learning_rate')
if self.is_group:
self.parameters = ParameterTuple(self.group_params)
self.weight_decay = tuple(self.group_weight_decay)
decay_filter = lambda x: x > 0
self.decay_flags = tuple(decay_filter(x) for x in self.weight_decay)
self.exec_weight_decay = any(self.decay_flags)
else:
self.parameters = ParameterTuple(parameters)
self.weight_decay = weight_decay * loss_scale
decay_filter = lambda x: 'beta' not in x.name and 'gamma' not in x.name
self.decay_flags = tuple(decay_filter(x) for x in self.parameters)
self.exec_weight_decay = self.weight_decay > 0
ps_filter = lambda x: x.is_param_ps
self.ps_parameters = tuple(ps_filter(x) for x in self.parameters)
self.reciprocal_scale = 1.0 / loss_scale
self.param_length = len(self.parameters)
self.map_ = C.Map()
use_parallel = context.get_auto_parallel_context("enable_parallel_optimizer")
self.use_parallel = use_parallel
if use_parallel:
if self.cls_name not in ["Lamb", "AdamWeightDecay"]:
raise RuntimeError("Optimizer segmentation does not support optimizer {}".format(self.cls_name))
if _get_parallel_mode() != ParallelMode.DATA_PARALLEL:
raise RuntimeError("Optimizer segmentation does not support parallel mode {}".format
(_get_parallel_mode()))
self.dev_num = _get_device_num()
if self.dev_num > self.param_length:
raise RuntimeError("Optimizer segmentation can not be applied when the number of parameters {} is"
" less than the number of devices {}".format(self.param_length, self.dev_num))
self.param_rank = self._get_parameter_group_id()
self.optim_filter = tuple(map(lambda x: x == _get_global_rank(), self.param_rank))
self.param_names = []
for param in self.parameters:
self.param_names.append(param.name)
else:
self.optim_filter = (True,) * self.param_length
T.not_: multitype_ops.logical_not,
T.eq: multitype_ops.equal,
T.ne: multitype_ops.not_equal,
T.lt: multitype_ops.less,
T.gt: multitype_ops.greater,
T.le: multitype_ops.less_equal,
T.ge: multitype_ops.greater_equal,
T.is_: F.is_,
T.is_not: F.is_not,
T.contains: multitype_ops.in_,
T.not_contains: multitype_ops.not_in_,
# system function
T.len: M.ms_len,
T.bool_: M.bool_,
T.map: C.Map(),
T.partial: F.partial,
T.zip: C.zip_operation,
T.print: F.print_,
T.enumerate: M.enumerate_,
T.isinstance: M.isinstance_,
# custom define operation
T.iter: M.ms_iter,
T.next: M.ms_next,
T.hasnext: M.hasnext,
T.make_tuple: F.make_tuple,
T.make_dict: F.make_dict,
T.make_list: F.make_list,
T.make_slice: F.make_slice,
def __init__(self):
super().__init__()
self.relu = nn.ReLU()
self.hypermap = C.Map()
def __init__(self, *args, **kwargs):
super(KungFuMomentum, self).__init__(*args, **kwargs)
from kungfu_mindspore_ops import KungFuAllReduce
self.map_ = C.Map()
self.all_reduce = KungFuAllReduce()
def __init__(self, learning_rate, parameters, weight_decay=0.0, loss_scale=1.0):
super(Optimizer, self).__init__(auto_prefix=False)
if parameters and not isinstance(parameters, list):
parameters = list(parameters)
if not parameters:
raise ValueError("Optimizer got an empty parameter list.")
if not isinstance(parameters[0], (dict, Parameter)):
raise TypeError("Only a list of Parameter or dict can be supported.")
if isinstance(loss_scale, int):
loss_scale = float(loss_scale)
validator.check_value_type("loss_scale", loss_scale, [float], self.cls_name)
validator.check_number_range("loss_scale", loss_scale, 0.0, float("inf"), Rel.INC_NEITHER, self.cls_name)
if isinstance(weight_decay, int):
weight_decay = float(weight_decay)
validator.check_value_type("weight_decay", weight_decay, [float], self.cls_name)
validator.check_number_range("weight_decay", weight_decay, 0.0, float("inf"), Rel.INC_LEFT, self.cls_name)
self.is_group = False
self.is_group_lr = False
self.is_group_params_ordered = False
self.loss_scale = loss_scale
if isinstance(learning_rate, int):
learning_rate = float(learning_rate)
if isinstance(learning_rate, float):
self.dynamic_lr = False
self.gather = None
self.assignadd = None
self.global_step = None
self.scalar_lr = learning_rate
else:
self.dynamic_lr = True
self.gather = P.GatherV2()
self.assignadd = P.AssignAdd()
self.global_step = Parameter(initializer(0, [1], mindspore.int32), name='global_step')
self.scalar_lr = None
learning_rate = self._get_single_lr(learning_rate)
if isinstance(parameters[0], dict):
self.is_group = True
self.group_params = []
self.group_lr = []
self.group_weight_decay = []
self._init_group_params(parameters, learning_rate, weight_decay)
if self.is_group_lr:
self.learning_rate = ParameterTuple(self.group_lr)
else:
self.learning_rate = Parameter(learning_rate, name="learning_rate")
if self.is_group:
self.parameters = ParameterTuple(self.group_params)
self.weight_decay = tuple(self.group_weight_decay)
decay_filter = lambda x: x > 0
self.decay_flags = tuple(decay_filter(x) for x in self.weight_decay)
else:
self.parameters = ParameterTuple(parameters)
self.weight_decay = weight_decay * loss_scale
decay_filter = lambda x: 'beta' not in x.name and 'gamma' not in x.name
self.decay_flags = tuple(decay_filter(x) for x in self.parameters)
self.reciprocal_scale = 1.0 / loss_scale
self.exec_weight_decay = any(self.decay_flags)
self.param_length = len(self.parameters)
self.map_ = C.Map()