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("start_learning_rate", start_learning_rate, [float], prim_name) validator.check_number_range("start_learning_rate rate", start_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) validator.check_integer('warmup_steps', warmup_steps, 0, Rel.GE, 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 _check_inputs(learning_rate, decay_rate, total_step, step_per_epoch, decay_epoch, is_stair): 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_float_positive('learning_rate', learning_rate, None) validator.check_float_legal_value('learning_rate', learning_rate, None) validator.check_float_positive('decay_rate', decay_rate, None) validator.check_float_legal_value('decay_rate', decay_rate, None) validator.check_value_type('is_stair', is_stair, [bool], None)
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 _check_inputs(learning_rate, decay_rate, total_step, step_per_epoch, decay_epoch, is_stair): validator.check_positive_int(total_step, 'total_step') validator.check_positive_int(step_per_epoch, 'step_per_epoch') validator.check_positive_int(decay_epoch, 'decay_epoch') validator.check_float_positive('learning_rate', learning_rate, None) validator.check_float_legal_value('learning_rate', learning_rate, None) validator.check_float_positive('decay_rate', decay_rate, None) validator.check_float_legal_value('decay_rate', decay_rate, None) validator.check_value_type('is_stair', is_stair, [bool], None)
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 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 piecewise_constant_lr(milestone, learning_rates): r""" Get piecewise constant learning rate. Calculate learning rate by given `milestone` and `learning_rates`. Let the value of `milestone` be :math:`(M_1, M_2, ..., M_N)` and the value of `learning_rates` be :math:`(x_1, x_2, ..., x_N)`. N is the length of `milestone`. Let the output learning rate be `y`. .. math:: y[i] = x_t,\ for\ i \in [M_{t-1}, M_t) Args: milestone (Union[list[int], tuple[int]]): A list of milestone. This list is a monotone increasing list. Every element is a milestone step, and must be greater than 0. learning_rates (Union[list[float], tuple[float]]): A list of learning rates. Returns: list[float]. The size of list is :math:`M_N`. Examples: >>> milestone = [2, 5, 10] >>> learning_rates = [0.1, 0.05, 0.01] >>> piecewise_constant_lr(milestone, learning_rates) [0.1, 0.1, 0.05, 0.05, 0.05, 0.01, 0.01, 0.01, 0.01, 0.01] """ validator.check_value_type('milestone', milestone, (tuple, list), None) validator.check_value_type('learning_rates', learning_rates, (tuple, list), None) if len(milestone) != len(learning_rates): raise ValueError( 'The size of `milestone` must be same with the size of `learning_rates`.' ) lr = [] last_item = 0 for i, item in enumerate(milestone): validator.check_integer(f'milestone[{i}]', item, 0, Rel.GT, None) validator.check_float_legal_value(f'learning_rates[{i}]', learning_rates[i], None) if item < last_item: raise ValueError( f'The value of milestone[{i}] must be greater than milestone[{i - 1}]' ) lr += [learning_rates[i]] * (item - last_item) last_item = item return lr
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})`, and :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) validator.check_float_positive('end_learning_rate', end_learning_rate, None) validator.check_float_legal_value('end_learning_rate', end_learning_rate, 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
def _check_learning_rate_value(learning_rate, end_learning_rate, decay_steps, power, prim_name): """Check the type of inputs.""" validator.check_float_positive('learning_rate', learning_rate, prim_name) validator.check_float_legal_value('learning_rate', 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)
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 __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_float_legal_value('weight_decay', weight_decay, None) if isinstance(loss_scale, int): loss_scale = float(loss_scale) validator.check_float_legal_value('loss_scale', loss_scale, None) if loss_scale <= 0.0: raise ValueError( "Loss scale should be greater than 0, but got {}".format( loss_scale)) self.loss_scale = loss_scale if weight_decay < 0.0: raise ValueError( "Weight decay should be equal or greater than 0, but got {}". format(weight_decay)) 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.")