def __init__(self, is_grad=True, sparse=False, reduction=None, smooth_factor=0, num_classes=2): super(SoftmaxCrossEntropyWithLogits, self).__init__(reduction) self.is_grad = is_grad self.sparse = sparse validator.check_integer("num_classes", num_classes, 1, Rel.GT, self.cls_name) validator.check_number_range("smooth_factor", smooth_factor, 0, 1, Rel.INC_BOTH, self.cls_name) self.smooth_factor = smooth_factor self.num_classes = num_classes self.softmax_cross_entropy = P.SoftmaxCrossEntropyWithLogits() self.one_hot = P.OneHot() self.on_value = Tensor(1.0 - self.smooth_factor, mstype.float32) self.off_value = Tensor( 1.0 * self.smooth_factor / (self.num_classes - 1), mstype.float32) self.is_cpugpu = context.get_context('device_target') in ["CPU", "GPU"] if self.is_cpugpu: self.sparse_softmax_cross_entropy = P.SparseSoftmaxCrossEntropyWithLogits( is_grad=self.is_grad)
def _preprocess_single_lr(self, learning_rate): """Check lr value, and convert lr to a float, a Tensor or a LearningRateSchedule.""" if isinstance(learning_rate, (float, int)): learning_rate = float(learning_rate) validator.check_number_range("learning rate", learning_rate, 0.0, float("inf"), Rel.INC_LEFT, self.cls_name) return learning_rate if isinstance(learning_rate, Tensor) and learning_rate.dim() == 0: return learning_rate self.dynamic_lr = True if isinstance(learning_rate, Iterable): return Tensor(np.array(list(learning_rate)).astype(np.float32)) if isinstance(learning_rate, Tensor): if learning_rate.dim() > 1: raise ValueError( "The dim of `Tensor` type Learning rate should be a 0 or 1," f"but got {learning_rate.dim()}.") if learning_rate.dim() == 1 and learning_rate.size() < 2: logger.warning( "If use `Tensor` type dynamic learning rate, please make sure that the number" "of elements in the tensor passed is greater than 1.") return learning_rate if isinstance(learning_rate, LearningRateSchedule): return learning_rate raise TypeError( "Learning rate should be int, float, Tensor, Iterable or LearningRateSchedule." )
def _preprocess_weight_decay(self, weight_decay): """Check weight decay, and convert int to float.""" if isinstance(weight_decay, (float, int)): weight_decay = float(weight_decay) validator.check_number_range("weight_decay", weight_decay, 0.0, float("inf"), Rel.INC_LEFT, self.cls_name) return weight_decay raise TypeError("Weight decay should be int or float.")
def __init__(self, params, learning_rate=1e-3, beta1=0.9, beta2=0.999, eps=1e-8, use_locking=False, use_nesterov=False, weight_decay=0.0, loss_scale=1.0): super(Adam, self).__init__(learning_rate, params, weight_decay, loss_scale) _check_param_value(beta1, beta2, eps, weight_decay, self.cls_name) validator.check_value_type("use_locking", use_locking, [bool], self.cls_name) validator.check_value_type("use_nesterov", use_nesterov, [bool], self.cls_name) validator.check_value_type("loss_scale", loss_scale, [float], self.cls_name) validator.check_number_range("loss_scale", loss_scale, 1.0, float("inf"), Rel.INC_LEFT, self.cls_name) self.beta1 = Tensor(beta1, mstype.float32) self.beta2 = Tensor(beta2, mstype.float32) self.beta1_power = Parameter(initializer(1, [1], mstype.float32), name="beta1_power") self.beta2_power = Parameter(initializer(1, [1], mstype.float32), name="beta2_power") self.eps = eps self.moment1 = self.parameters.clone(prefix="moment1", init='zeros') self.moment2 = self.parameters.clone(prefix="moment2", init='zeros') self.hyper_map = C.HyperMap() self.opt = P.Adam(use_locking, use_nesterov) self.pow = P.Pow() self.sqrt = P.Sqrt() self.one = Tensor(np.array([1.0]).astype(np.float32)) self.realdiv = P.RealDiv()
def _check_param_value(decay_steps, warmup_steps, start_learning_rate, end_learning_rate, power, beta1, beta2, eps, weight_decay, prim_name): """Check the type of inputs.""" validator.check_value_type("decay_steps", decay_steps, [int], prim_name) validator.check_value_type("warmup_steps", warmup_steps, [int], prim_name) validator.check_value_type("start_learning_rate", start_learning_rate, [float], prim_name) validator.check_value_type("end_learning_rate", end_learning_rate, [float], prim_name) validator.check_value_type("power", power, [float], prim_name) validator.check_value_type("beta1", beta1, [float], prim_name) validator.check_value_type("beta2", beta2, [float], prim_name) validator.check_value_type("eps", eps, [float], prim_name) validator.check_value_type("weight_dacay", weight_decay, [float], prim_name) validator.check_number_range("decay_steps", decay_steps, 1, float("inf"), Rel.INC_LEFT, prim_name) validator.check_number_range("beta1", beta1, 0.0, 1.0, Rel.INC_NEITHER, prim_name) validator.check_number_range("beta2", beta2, 0.0, 1.0, Rel.INC_NEITHER, prim_name) validator.check_number_range("eps", eps, 0.0, float("inf"), Rel.INC_NEITHER, prim_name) validator.check_number_range("weight_decay", weight_decay, 0.0, float("inf"), Rel.INC_LEFT, prim_name)
def _check_learning_rate_value(learning_rate, end_learning_rate, decay_steps, power, prim_name): """Check the type of inputs.""" validator.check_value_type("learning_rate", learning_rate, [float], prim_name) validator.check_number_range("learning_rate", learning_rate, 0.0, float("inf"), Rel.INC_LEFT, prim_name) validator.check_value_type("end_learning_rate", end_learning_rate, [float], prim_name) validator.check_number_range("end_learning_rate", end_learning_rate, 0.0, float("inf"), Rel.INC_LEFT, prim_name) validator.check_float_positive('power', power, prim_name) validator.check_float_legal_value('power', power, prim_name) validator.check_integer('decay_steps', decay_steps, 0, Rel.GT, prim_name)
def __init__(self, max_val=1.0, filter_size=11, filter_sigma=1.5, k1=0.01, k2=0.03): super(SSIM, self).__init__() validator.check_value_type('max_val', max_val, [int, float], self.cls_name) validator.check_number('max_val', max_val, 0.0, Rel.GT, self.cls_name) self.max_val = max_val self.filter_size = validator.check_integer('filter_size', filter_size, 1, Rel.GE, self.cls_name) self.filter_sigma = validator.check_float_positive('filter_sigma', filter_sigma, self.cls_name) validator.check_value_type('k1', k1, [float], self.cls_name) self.k1 = validator.check_number_range('k1', k1, 0.0, 1.0, Rel.INC_NEITHER, self.cls_name) validator.check_value_type('k2', k2, [float], self.cls_name) self.k2 = validator.check_number_range('k2', k2, 0.0, 1.0, Rel.INC_NEITHER, self.cls_name) self.mean = P.DepthwiseConv2dNative(channel_multiplier=1, kernel_size=filter_size)
def _init_group_params(self, parameters, learning_rate, weight_decay): """Init learning rate or weight decay in group params.""" origin_dynamic_lr = self.dynamic_lr self._parse_group_params(parameters, learning_rate) if self.dynamic_lr and not origin_dynamic_lr: self.gather = P.GatherV2() self.assignadd = P.AssignAdd() self.global_step = Parameter(initializer(0, [1], mindspore.int32), name='global_step') params_store = [] for group_param in parameters: if 'order_params' in group_param.keys(): ordered_parameters = group_param['order_params'] continue self.group_params += group_param['params'] if 'lr' in group_param.keys(): params_dynamic_lr = isinstance(group_param['lr'], (Iterable, Tensor)) if self.dynamic_lr and not params_dynamic_lr: lr = Tensor(np.array([group_param['lr']] * self.dynamic_lr_length).astype(np.float32)) else: lr = self._get_single_lr(group_param['lr']) else: if self.dynamic_lr and not origin_dynamic_lr: lr = Tensor(np.array([self.scalar_lr] * self.dynamic_lr_length).astype(np.float32)) else: lr = learning_rate if 'weight_decay' in group_param.keys(): validator.check_float_legal_value('weight_decay', group_param['weight_decay'], None) validator.check_number_range('weight_decay', group_param['weight_decay'], 0.0, float("inf"), Rel.INC_LEFT, self.cls_name) weight_decay_ = group_param['weight_decay'] * self.loss_scale else: weight_decay_ = weight_decay * self.loss_scale for key in group_param.keys(): if key not in ('params', 'lr', 'weight_decay'): logger.warning(f"The optimizer cannot parse '{key}' when setting parameter groups.") for param in group_param['params']: validator.check_value_type("parameter", param, [Parameter], self.cls_name) if param.name in params_store: raise RuntimeError(f"The {param.name} parameter has appeared in parameter groups.") params_store.append(param.name) self.group_lr.append(Parameter(lr, name="lr_" + param.name)) self.group_weight_decay.append(weight_decay_) if self.is_group_params_ordered: self._order_and_adjust_group_params(ordered_parameters, learning_rate, weight_decay)
def _check_param_value(decay_steps, warmup_steps, start_learning_rate, end_learning_rate, power, beta1, beta2, eps, weight_decay, prim_name): """Check the type of inputs.""" _ = warmup_steps validator.check_float_positive('start_learning_rate', start_learning_rate, prim_name) validator.check_float_legal_value('start_learning_rate', start_learning_rate, prim_name) validator.check_float_positive('end_learning_rate', end_learning_rate, prim_name) validator.check_float_legal_value('end_learning_rate', end_learning_rate, prim_name) validator.check_float_positive('power', power, prim_name) validator.check_float_legal_value('power', power, prim_name) validator.check_integer('decay_steps', decay_steps, 0, Rel.GT, prim_name) validator.check_integer('warmup_steps', decay_steps, 0, Rel.GT, prim_name) validator.check_value_type("beta1", beta1, [float], prim_name) validator.check_value_type("beta2", beta2, [float], prim_name) validator.check_value_type("eps", eps, [float], prim_name) validator.check_value_type("weight_dacay", weight_decay, [float], prim_name) validator.check_number_range("beta1", beta1, 0.0, 1.0, Rel.INC_NEITHER, prim_name) validator.check_number_range("beta2", beta2, 0.0, 1.0, Rel.INC_NEITHER, prim_name) validator.check_number_range("eps", eps, 0.0, float("inf"), Rel.INC_NEITHER, prim_name) validator.check_number_range("weight_decay", weight_decay, 0.0, float("inf"), Rel.INC_LEFT, prim_name)
def cosine_decay_lr(min_lr, max_lr, total_step, step_per_epoch, decay_epoch): r""" Calculate learning rate base on cosine decay function. For the i-th step, the formula of computing decayed_learning_rate[i] is: .. math:: decayed\_learning\_rate[i] = min\_learning\_rate + 0.5 * (max\_learning\_rate - min\_learning\_rate) * (1 + cos(\frac{current\_epoch}{decay\_epoch}\pi)) Where :math:`current\_epoch=floor(\frac{i}{step\_per\_epoch})`. Args: min_lr (float): The minimum value of learning rate. max_lr (float): The maximum value of learning rate. total_step (int): The total number of steps. step_per_epoch (int): The number of steps in per epoch. decay_epoch (int): A value used to calculate decayed learning rate. Returns: list[float]. The size of list is `total_step`. Examples: >>> min_lr = 0.01 >>> max_lr = 0.1 >>> total_step = 6 >>> step_per_epoch = 2 >>> decay_epoch = 2 >>> cosine_decay_lr(min_lr, max_lr, total_step, step_per_epoch, decay_epoch) [0.1, 0.1, 0.05500000000000001, 0.05500000000000001, 0.01, 0.01] """ if not isinstance(min_lr, float): raise TypeError("min_lr must be float.") validator.check_number_range("min_lr", min_lr, 0.0, float("inf"), Rel.INC_LEFT, None) validator.check_float_positive('max_lr', max_lr, None) validator.check_float_legal_value('max_lr', max_lr, None) validator.check_integer('total_step', total_step, 0, Rel.GT, None) validator.check_integer('step_per_epoch', step_per_epoch, 0, Rel.GT, None) validator.check_integer('decay_epoch', decay_epoch, 0, Rel.GT, None) if min_lr >= max_lr: raise ValueError('`max_lr` should be greater than `min_lr`.') delta = 0.5 * (max_lr - min_lr) lr = [] for i in range(total_step): tmp_epoch = min(math.floor(i / step_per_epoch), decay_epoch) lr.append(min_lr + delta * (1 + math.cos(math.pi * tmp_epoch / decay_epoch))) return lr
def __init__(self, margin=0.0, reduction="mean"): super(CosineEmbeddingLoss, self).__init__(reduction) self.reduce_sum = P.ReduceSum() self.maximum = P.Maximum() validator.check_value_type("margin", margin, [float], self.cls_name) self.margin = validator.check_number_range("margin", margin, -1.0, 1.0, Rel.INC_BOTH, self.cls_name)
def __init__(self, central_fraction): super(CentralCrop, self).__init__() validator.check_value_type("central_fraction", central_fraction, [float], self.cls_name) self.central_fraction = validator.check_number_range('central_fraction', central_fraction, 0.0, 1.0, Rel.INC_RIGHT, self.cls_name) self.central_fraction_tensor = Tensor(np.array([central_fraction]).astype(np.float64)) self.slice = P.Slice()
def _check_param_value(beta1, beta2, eps, weight_decay, prim_name): """Check the type of inputs.""" validator.check_value_type("beta1", beta1, [float], prim_name) validator.check_value_type("beta2", beta2, [float], prim_name) validator.check_value_type("eps", eps, [float], prim_name) validator.check_value_type("weight_dacay", weight_decay, [float], prim_name) validator.check_number_range("beta1", beta1, 0.0, 1.0, Rel.INC_NEITHER, prim_name) validator.check_number_range("beta2", beta2, 0.0, 1.0, Rel.INC_NEITHER, prim_name) validator.check_number_range("eps", eps, 0.0, float("inf"), Rel.INC_NEITHER, prim_name) validator.check_number_range("weight_decay", weight_decay, 0.0, float("inf"), Rel.INC_LEFT, prim_name)
def warmup_lr(learning_rate, total_step, step_per_epoch, warmup_epoch): r""" Get learning rate warming up. For the i-th step, the formula of computing warmup_learning_rate[i] is: .. math:: warmup\_learning\_rate[i] = learning\_rate * tmp\_epoch / tmp\_warmup\_epoch Where :math:`tmp\_epoch=min(current\_epoch, warmup\_epoch),\ current\_epoch=floor(\frac{i}{step\_per\_epoch})` Args: learning_rate (float): The initial value of learning rate. warmup_steps (int): The warm up steps of learning rate. Inputs: Tensor. The current step number. Returns: Tensor. The learning rate value for the current step. Examples: >>> learning_rate = 0.1 >>> total_step = 6 >>> step_per_epoch = 2 >>> warmup_epoch = 2 >>> warmup_lr(learning_rate, total_step, step_per_epoch, warmup_epoch) [0.0, 0.0, 0.05, 0.05, 0.1, 0.1] """ if not isinstance(learning_rate, float): raise TypeError("learning_rate must be float.") validator.check_number_range("learning_rate", learning_rate, 0.0, float("inf"), Rel.INC_LEFT, None) validator.check_integer('warmup_epoch', warmup_epoch, 0, Rel.GT, None) validator.check_integer('total_step', total_step, 0, Rel.GT, None) validator.check_integer('step_per_epoch', step_per_epoch, 0, Rel.GT, None) function = lambda x, y: (x, min(x, y)) lr = [] for i in range(total_step): current_epoch = math.floor(i / step_per_epoch) warmup_epoch, tmp_epoch = function(warmup_epoch, current_epoch) lr.append(learning_rate * tmp_epoch / warmup_epoch) return lr
def _get_single_lr(self, learning_rate): """Get learning rate in Tensor type.""" if isinstance(learning_rate, float): validator.check_number_range("learning rate", learning_rate, 0.0, float("inf"), Rel.INC_LEFT, self.cls_name) lr = Tensor(learning_rate, mstype.float32) elif isinstance(learning_rate, Iterable): lr = Tensor(np.array(list(learning_rate)).astype(np.float32)) elif isinstance(learning_rate, Tensor): if learning_rate.dim() > 1: raise ValueError("Learning rate should be a 0 or 1 dim `Tensor`," f"but got {learning_rate.dim()}.") if learning_rate.dim() == 1 and learning_rate.size() < 2: logger.warning("If want to use the dynamic learning rate, please make sure that the number " "of elements in the list, tuple or tensor passed is greater than 1.") lr = learning_rate else: raise TypeError("Learning rate should be float, Tensor or Iterable.") return lr
def __init__(self, params, learning_rate=0.1, decay=0.9, momentum=0.0, epsilon=1e-10, use_locking=False, centered=False, loss_scale=1.0, weight_decay=0.0): super(RMSProp, self).__init__(learning_rate, params, weight_decay, loss_scale) validator.check_value_type("decay", decay, [float], self.cls_name) validator.check_number_range("decay", decay, 0.0, float("inf"), Rel.INC_LEFT, self.cls_name) validator.check_value_type("momentum", momentum, [float], self.cls_name) validator.check_number_range("momentum", momentum, 0.0, float("inf"), Rel.INC_LEFT, self.cls_name) validator.check_value_type("epsilon", epsilon, [float], self.cls_name) validator.check_number_range("epsilon", epsilon, 0.0, float("inf"), Rel.INC_NEITHER, self.cls_name) validator.check_value_type("use_locking", use_locking, [bool], self.cls_name) validator.check_value_type("centered", centered, [bool], self.cls_name) self.centered = centered if centered: self.opt = P.ApplyCenteredRMSProp(use_locking) self.mg = self.parameters.clone(prefix="mean_grad", init='zeros') else: self.opt = P.ApplyRMSProp(use_locking) self.momentum = momentum self.ms = self.parameters.clone(prefix="mean_square", init='zeros') self.moment = self.parameters.clone(prefix="moment", init='zeros') self.hyper_map = C.HyperMap() self.epsilon = epsilon self.decay = decay
def _check_param_value(accum, l1, l2, use_locking, prim_name=None): """Check inputs param.""" validator.check_value_type("accum", accum, [float], prim_name) validator.check_value_type("l1", l1, [float], prim_name) validator.check_value_type("l2", l2, [float], prim_name) validator.check_value_type("use_locking", use_locking, [bool], prim_name) validator.check_number_range("accum", accum, 0.0, float("inf"), Rel.INC_LEFT, prim_name) validator.check_number_range("l1", l1, 0.0, float("inf"), Rel.INC_LEFT, prim_name) validator.check_number_range("l2", l2, 0.0, float("inf"), Rel.INC_LEFT, prim_name)
def _check_param_value(beta1, beta2, eps, prim_name): validator.check_value_type("beta1", beta1, [float], prim_name) validator.check_value_type("beta2", beta2, [float], prim_name) validator.check_value_type("eps", eps, [float], prim_name) validator.check_number_range("beta1", beta1, 0.0, 1.0, Rel.INC_NEITHER, prim_name) validator.check_number_range("beta2", beta2, 0.0, 1.0, Rel.INC_NEITHER, prim_name) validator.check_number_range("eps", eps, 0.0, float("inf"), Rel.INC_NEITHER, prim_name)
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], None) validator.check_number_range("loss_scale", loss_scale, 0.0, float("inf"), Rel.INC_NEITHER, None) if isinstance(weight_decay, int): weight_decay = float(weight_decay) validator.check_value_type("weight_decay", weight_decay, [float], None) validator.check_number_range("weight_decay", weight_decay, 0.0, float("inf"), Rel.INC_LEFT, None) self.is_group = False self.is_group_lr = False self.loss_scale = loss_scale 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.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.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)
def __init__(self, learning_rate, parameters, weight_decay=0.0, loss_scale=1.0, decay_filter=lambda x: 'beta' not in x.name and 'gamma' not in x.name): super(Optimizer, self).__init__(auto_prefix=False) if isinstance(learning_rate, float): self.dynamic_lr = False self.gather = None self.assignadd = None self.global_step = None validator.check_number_range("learning rate", learning_rate, 0.0, float("inf"), Rel.INC_LEFT, self.cls_name) learning_rate = Tensor(learning_rate, mstype.float32) 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') if isinstance(learning_rate, Iterable): learning_rate = Tensor( np.array(list(learning_rate)).astype(np.float32)) elif isinstance(learning_rate, Tensor): if learning_rate.dim() > 1: raise ValueError( "Learning rate should be a 0 or 1 dim `Tensor`," f"but got {learning_rate.dim()}.") if learning_rate.dim() == 1 and learning_rate.size() < 2: logger.warning( "If want to use the dynamic learning rate, please make sure that the number " "of elements in the list, tuple or tensor passed is greater than 1." ) else: raise TypeError( "Learning rate should be float, Tensor or Iterable.") if isinstance(weight_decay, int): weight_decay = float(weight_decay) validator.check_value_type("weight_decay", weight_decay, [float], None) validator.check_number_range("weight_decay", weight_decay, 0.0, float("inf"), Rel.INC_LEFT, None) if isinstance(loss_scale, int): loss_scale = float(loss_scale) validator.check_value_type("loss_scale", loss_scale, [float], None) validator.check_number_range("loss_scale", loss_scale, 0.0, float("inf"), Rel.INC_NEITHER, None) self.loss_scale = loss_scale self.learning_rate = Parameter(learning_rate, name="learning_rate") self.parameters = ParameterTuple(parameters) self.reciprocal_scale = 1.0 / loss_scale self.weight_decay = weight_decay * loss_scale self.decay_flags = tuple(decay_filter(x) for x in self.parameters) if not self.parameters: raise ValueError("optimizer got an empty parameter list.")
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
def _init_group_params(self, parameters, learning_rate, weight_decay): """Init learning rate or weight decay in group params.""" origin_dynamic_lr = self.dynamic_lr if self.dynamic_lr: dynamic_lr_length = learning_rate.size() else: dynamic_lr_length = 0 for group_param in parameters: lr_length = dynamic_lr_length if 'lr' in group_param.keys(): self._get_single_lr(group_param['lr']) if isinstance(group_param['lr'], Iterable): lr_length = len(group_param['lr']) self.dynamic_lr = True elif isinstance(group_param['lr'], Tensor): lr_length = group_param['lr'].size() self.dynamic_lr = True if dynamic_lr_length not in (lr_length, 0): raise ValueError( "The dynamic learning rate in group should be the same size." ) dynamic_lr_length = lr_length if self.dynamic_lr and not origin_dynamic_lr: self.gather = P.GatherV2() self.assignadd = P.AssignAdd() self.global_step = Parameter(initializer(0, [1], mindspore.int32), name='global_step') params_store = [] for group_param in parameters: self.params += group_param['params'] if 'lr' in group_param.keys(): params_dynamic_lr = isinstance(group_param['lr'], (Iterable, Tensor)) if self.dynamic_lr and not params_dynamic_lr: lr = Tensor( np.array([group_param['lr']] * dynamic_lr_length).astype(np.float32)) else: lr = self._get_single_lr(group_param['lr']) else: if self.dynamic_lr and not origin_dynamic_lr: lr = Tensor( np.array([self.scalar_lr] * dynamic_lr_length).astype( np.float32)) else: lr = learning_rate if 'weight_decay' in group_param.keys(): validator.check_float_legal_value('weight_decay', group_param['weight_decay'], None) validator.check_number_range('weight_decay', group_param['weight_decay'], 0.0, float("inf"), Rel.INC_LEFT, self.cls_name) weight_decay_ = group_param['weight_decay'] * self.loss_scale else: weight_decay_ = weight_decay * self.loss_scale for param in group_param['params']: if param in params_store: raise RuntimeError( f"The {param.name} parameter has appeared in parameter groups." ) params_store.append(param) self.group_lr.append(Parameter(lr, name="lr_" + param.name)) self.group_weight_decay.append(weight_decay_)
def polynomial_decay_lr(learning_rate, end_learning_rate, total_step, step_per_epoch, decay_epoch, power, update_decay_epoch=False): r""" Calculate learning rate base on polynomial decay function. For the i-th step, the formula of computing decayed_learning_rate[i] is: .. math:: decayed\_learning\_rate[i] = (learning\_rate - end\_learning\_rate) * (1 - tmp\_epoch / tmp\_decay\_epoch)^{power} + end\_learning\_rate Where: .. math:: `tmp\_epoch = min(current\_epoch, decay\_epoch), current\_epoch=floor(\frac{i}{step\_per\_epoch})`, .. math:: `tmp\_decay\_epoch = decay\_epoch`. If `update_decay_epoch` is true, update the value of `tmp_decay_epoch` every epoch. The formula is: .. math:: `tmp\_decay\_epoch = decay\_epoch * ceil(current\_epoch / decay\_epoch)` Args: learning_rate (float): The initial value of learning rate. end_learning_rate (float): The end value of learning rate. total_step (int): The total number of steps. step_per_epoch (int): The number of steps in per epoch. decay_epoch (int): A value used to calculate decayed learning rate. power (float): A value used to calculate decayed learning rate. This parameter should be greater than 0. update_decay_epoch (bool): If true, update `decay_epoch`. Default: False. Returns: list[float]. The size of list is `total_step`. Examples: >>> learning_rate = 0.1 >>> end_learning_rate = 0.01 >>> total_step = 6 >>> step_per_epoch = 2 >>> decay_epoch = 2 >>> power = 0.5 >>> polynomial_decay_lr(learning_rate, end_learning_rate, total_step, step_per_epoch, decay_epoch, power) [0.1, 0.1, 0.07363961030678928, 0.07363961030678928, 0.01, 0.01] """ validator.check_float_positive('learning_rate', learning_rate, None) validator.check_float_legal_value('learning_rate', learning_rate, None) if not isinstance(end_learning_rate, float): raise TypeError("end_learning_rate must be float.") validator.check_number_range("end_learning_rate", end_learning_rate, 0.0, float("inf"), Rel.INC_LEFT, None) validator.check_float_positive('power', power, None) validator.check_float_legal_value('power', power, None) validator.check_integer('total_step', total_step, 0, Rel.GT, None) validator.check_integer('step_per_epoch', step_per_epoch, 0, Rel.GT, None) validator.check_integer('decay_epoch', decay_epoch, 0, Rel.GT, None) validator.check_value_type('update_decay_epoch', update_decay_epoch, [bool], None) origin_decay_epoch = decay_epoch function = lambda x, y: (x, min(x, y)) if update_decay_epoch: function = lambda x, y: (origin_decay_epoch * max( math.ceil(y / origin_decay_epoch), 1), y) lr = [] delta = learning_rate - end_learning_rate for i in range(total_step): current_epoch = math.floor(i / step_per_epoch) decay_epoch, tmp_epoch = function(decay_epoch, current_epoch) lr.append(delta * (1 - tmp_epoch / decay_epoch)**power + end_learning_rate) return lr