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_non_negative_float(weight_decay, "weight_decay", self.cls_name) return weight_decay raise TypeError("Weight decay should be int or float.")
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_non_negative_float(learning_rate, "learning rate", self.cls_name) return learning_rate if isinstance(learning_rate, Tensor) and learning_rate.ndim == 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.ndim > 1: raise ValueError( "The dim of `Tensor` type Learning rate should be a 0 or 1," f"but got {learning_rate.ndim}.") if learning_rate.ndim == 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 __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_non_negative_float(decay, "decay", self.cls_name) validator.check_value_type("momentum", momentum, [float], self.cls_name) validator.check_non_negative_float(momentum, "momentum", self.cls_name) validator.check_value_type("epsilon", epsilon, [float], self.cls_name) validator.check_positive_float(epsilon, "epsilon", 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='ones') self.moment = self.parameters.clone(prefix="moment", init='zeros') self.hyper_map = C.HyperMap() self.epsilon = epsilon self.decay = decay
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_float_range(beta1, 0.0, 1.0, Rel.INC_NEITHER, "beta1", prim_name) validator.check_float_range(beta2, 0.0, 1.0, Rel.INC_NEITHER, "beta2", prim_name) validator.check_positive_float(eps, "eps", prim_name) validator.check_non_negative_float(weight_decay, "weight_decay", 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 >>> output = cosine_decay_lr(min_lr, max_lr, total_step, step_per_epoch, decay_epoch) >>> print(output) [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_non_negative_float(min_lr, "min_lr", None) validator.check_positive_float(max_lr, 'max_lr') validator.check_is_float(max_lr, 'max_lr') 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') 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 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 >>> output = warmup_lr(learning_rate, total_step, step_per_epoch, warmup_epoch) >>> print(output) [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_non_negative_float(learning_rate, "learning_rate", None) validator.check_positive_int(warmup_epoch, 'warmup_epoch') validator.check_positive_int(total_step, 'total_step') validator.check_positive_int(step_per_epoch, 'step_per_epoch') 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 _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_non_negative_float(accum, "accum", prim_name) validator.check_non_negative_float(l1, "l1", prim_name) validator.check_non_negative_float(l2, "l2", prim_name)
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) .. math:: 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 must 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 >>> r = polynomial_decay_lr(learning_rate, end_learning_rate, total_step, step_per_epoch, decay_epoch, power) >>> print(r) [0.1, 0.1, 0.07363961030678928, 0.07363961030678928, 0.01, 0.01] """ validator.check_positive_float(learning_rate, 'learning_rate') validator.check_is_float(learning_rate, 'learning_rate') if not isinstance(end_learning_rate, float): raise TypeError("end_learning_rate must be float.") validator.check_non_negative_float(end_learning_rate, "end_learning_rate", None) validator.check_positive_float(power, 'power') validator.check_is_float(power, 'power') 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_value_type('update_decay_epoch', update_decay_epoch, [bool]) 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_param_value(accum, update_slots, prim_name=None): """Check inputs param.""" validator.check_value_type("accum", accum, [float], prim_name) validator.check_value_type("update_slots", update_slots, [bool], prim_name) validator.check_non_negative_float(accum, "accum", prim_name)