o i e@sHdZddlZddlZddlmmZddlmZddl m Z ddl m Z ddlmZedGdd d Zed Gd d d eZed GdddeZedGdddeZedGdddeZeddGdddeZeddGdddeZGdddeZGd d!d!eZed"d)d$d%Zed&d*d'd(ZdS)+z)Various learning rate schedule functions.N)backend)serialization_lib) serialization) keras_exportz/keras.optimizers.schedules.LearningRateSchedulec@s8eZdZdZejddZejddZeddZ dS) LearningRateScheduleaThe learning rate schedule base class. You can use a learning rate schedule to modulate how the learning rate of your optimizer changes over time. Several built-in learning rate schedules are available, such as `tf.keras.optimizers.schedules.ExponentialDecay` or `tf.keras.optimizers.schedules.PiecewiseConstantDecay`: ```python lr_schedule = keras.optimizers.schedules.ExponentialDecay( initial_learning_rate=1e-2, decay_steps=10000, decay_rate=0.9) optimizer = keras.optimizers.SGD(learning_rate=lr_schedule) ``` A `LearningRateSchedule` instance can be passed in as the `learning_rate` argument of any optimizer. To implement your own schedule object, you should implement the `__call__` method, which takes a `step` argument (scalar integer tensor, the current training step count). Like for any other Keras object, you can also optionally make your object serializable by implementing the `get_config` and `from_config` methods. Example: ```python class MyLRSchedule(tf.keras.optimizers.schedules.LearningRateSchedule): def __init__(self, initial_learning_rate): self.initial_learning_rate = initial_learning_rate def __call__(self, step): return self.initial_learning_rate / (step + 1) optimizer = tf.keras.optimizers.SGD(learning_rate=MyLRSchedule(0.1)) ``` cCtd|jjd)NLearning rate schedule 'z'' must override `__call__(self, step)`.NotImplementedError __class____name__selfsteprd/var/www/myenv/lib/python3.10/site-packages/keras/src/optimizers/schedules/learning_rate_schedule.py__call__JzLearningRateSchedule.__call__cCr)Nrz;' must override `get_config()` in order to be serializable.r rrrr get_configQrzLearningRateSchedule.get_configcCs|di|S)zInstantiates a `LearningRateSchedule` from its config. Args: config: Output of `get_config()`. Returns: A `LearningRateSchedule` instance. Nrr)clsconfigrrr from_configXs z LearningRateSchedule.from_configN) r __module__ __qualname____doc__abcabstractmethodrr classmethodrrrrrrs*  rz+keras.optimizers.schedules.ExponentialDecayc6eZdZdZ  d fdd ZddZdd ZZS) ExponentialDecayaQA LearningRateSchedule that uses an exponential decay schedule. When training a model, it is often useful to lower the learning rate as the training progresses. This schedule applies an exponential decay function to an optimizer step, given a provided initial learning rate. The schedule is a 1-arg callable that produces a decayed learning rate when passed the current optimizer step. This can be useful for changing the learning rate value across different invocations of optimizer functions. It is computed as: ```python def decayed_learning_rate(step): return initial_learning_rate * decay_rate ^ (step / decay_steps) ``` If the argument `staircase` is `True`, then `step / decay_steps` is an integer division and the decayed learning rate follows a staircase function. You can pass this schedule directly into a `tf.keras.optimizers.Optimizer` as the learning rate. Example: When fitting a Keras model, decay every 100000 steps with a base of 0.96: ```python initial_learning_rate = 0.1 lr_schedule = tf.keras.optimizers.schedules.ExponentialDecay( initial_learning_rate, decay_steps=100000, decay_rate=0.96, staircase=True) model.compile(optimizer=tf.keras.optimizers.SGD(learning_rate=lr_schedule), loss='sparse_categorical_crossentropy', metrics=['accuracy']) model.fit(data, labels, epochs=5) ``` The learning rate schedule is also serializable and deserializable using `tf.keras.optimizers.schedules.serialize` and `tf.keras.optimizers.schedules.deserialize`. Returns: A 1-arg callable learning rate schedule that takes the current optimizer step and outputs the decayed learning rate, a scalar `Tensor` of the same type as `initial_learning_rate`. FNc,t||_||_||_||_||_dS)aApplies exponential decay to the learning rate. Args: initial_learning_rate: A scalar `float32` or `float64` `Tensor` or a Python number. The initial learning rate. decay_steps: A scalar `int32` or `int64` `Tensor` or a Python number. Must be positive. See the decay computation above. decay_rate: A scalar `float32` or `float64` `Tensor` or a Python number. The decay rate. staircase: Boolean. If `True` decay the learning rate at discrete intervals name: String. Optional name of the operation. Defaults to 'ExponentialDecay'. Nsuper__init__initial_learning_rate decay_steps decay_rate staircasenamerr%r&r'r(r)r rrr$s  zExponentialDecay.__init__c Cst|jpd?}tj|jdd}|j}t|j|}t|j|}t||}||}|j r4t |}tj |t |||dWdS1sJwYdS)Nr r%r)) tf name_scoper)convert_to_tensorr%dtypecastr&r'r(floormultiplypow) rrr)r%r0r&r'global_step_recompprrrrs  $zExponentialDecay.__call__cC|j|j|j|j|jdSNr%r&r'r(r)r9rrrrr zExponentialDecay.get_configFNr rrrr$rr __classcell__rrr+rr es7r z1keras.optimizers.schedules.PiecewiseConstantDecaycs2eZdZdZd fdd ZddZddZZS) PiecewiseConstantDecayaA LearningRateSchedule that uses a piecewise constant decay schedule. The function returns a 1-arg callable to compute the piecewise constant when passed the current optimizer step. This can be useful for changing the learning rate value across different invocations of optimizer functions. Example: use a learning rate that's 1.0 for the first 100001 steps, 0.5 for the next 10000 steps, and 0.1 for any additional steps. ```python step = tf.Variable(0, trainable=False) boundaries = [100000, 110000] values = [1.0, 0.5, 0.1] learning_rate_fn = keras.optimizers.schedules.PiecewiseConstantDecay( boundaries, values) # Later, whenever we perform an optimization step, we pass in the step. learning_rate = learning_rate_fn(step) ``` You can pass this schedule directly into a `tf.keras.optimizers.Optimizer` as the learning rate. The learning rate schedule is also serializable and deserializable using `tf.keras.optimizers.schedules.serialize` and `tf.keras.optimizers.schedules.deserialize`. Returns: A 1-arg callable learning rate schedule that takes the current optimizer step and outputs the decayed learning rate, a scalar `Tensor` of the same type as the boundary tensors. The output of the 1-arg function that takes the `step` is `values[0]` when `step <= boundaries[0]`, `values[1]` when `step > boundaries[0]` and `step <= boundaries[1]`, ..., and values[-1] when `step > boundaries[-1]`. Nc s^tt|t|dkr$td|dt|d|dt|d ||_||_||_dS)aPiecewise constant from boundaries and interval values. Args: boundaries: A list of `Tensor`s or `int`s or `float`s with strictly increasing entries, and with all elements having the same type as the optimizer step. values: A list of `Tensor`s or `float`s or `int`s that specifies the values for the intervals defined by `boundaries`. It should have one more element than `boundaries`, and all elements should have the same type. name: A string. Optional name of the operation. Defaults to 'PiecewiseConstant'. Raises: ValueError: if the number of elements in the lists do not match. zZThe length of boundaries should be 1 less than the length of values. Received: boundaries=z of length z , and values=.N)r#r$len ValueError boundariesvaluesr))rrCrDr)r+rrr$s  zPiecewiseConstantDecay.__init__c s^t|jpdtjtjtj|j}tjtjtj|jt|}t |D]\}}|j j |j j krDt ||j j }|||<q,g}| ||dkfddf| ||dkfddft|dd|ddddD]\}}} ||k||k@} | | | fddfqwfd d} tj|| d d WdS1swYdS) NPiecewiseConstantrcdSNrrrrDrr%z1PiecewiseConstantDecay.__call__..crF)NrKrrrHrrrI'rJr?cSs|SNr)vrrrrI.scrFrGrrrHrrrI2rJT) exclusive)r-r.r)nest map_structurer/flattenrCrD enumerater0 base_dtyper1appendzipcase) rrrCx_recompib pred_fn_pairslowhighrMpreddefaultrrHrrs4  $zPiecewiseConstantDecay.__call__cCs|j|j|jdS)NrCrDr)r_rrrrr5sz!PiecewiseConstantDecay.get_configrLr<rrr+rr>s $r>z*keras.optimizers.schedules.PolynomialDecayc:eZdZdZ    d fdd Zdd Zd d ZZS) PolynomialDecaya A LearningRateSchedule that uses a polynomial decay schedule. It is commonly observed that a monotonically decreasing learning rate, whose degree of change is carefully chosen, results in a better performing model. This schedule applies a polynomial decay function to an optimizer step, given a provided `initial_learning_rate`, to reach an `end_learning_rate` in the given `decay_steps`. It requires a `step` value to compute the decayed learning rate. You can just pass a TensorFlow variable that you increment at each training step. The schedule is a 1-arg callable that produces a decayed learning rate when passed the current optimizer step. This can be useful for changing the learning rate value across different invocations of optimizer functions. It is computed as: ```python def decayed_learning_rate(step): step = min(step, decay_steps) return ((initial_learning_rate - end_learning_rate) * (1 - step / decay_steps) ^ (power) ) + end_learning_rate ``` If `cycle` is True then a multiple of `decay_steps` is used, the first one that is bigger than `step`. ```python def decayed_learning_rate(step): decay_steps = decay_steps * ceil(step / decay_steps) return ((initial_learning_rate - end_learning_rate) * (1 - step / decay_steps) ^ (power) ) + end_learning_rate ``` You can pass this schedule directly into a `tf.keras.optimizers.Optimizer` as the learning rate. Example: Fit a model while decaying from 0.1 to 0.01 in 10000 steps using sqrt (i.e. power=0.5): ```python ... starter_learning_rate = 0.1 end_learning_rate = 0.01 decay_steps = 10000 learning_rate_fn = tf.keras.optimizers.schedules.PolynomialDecay( starter_learning_rate, decay_steps, end_learning_rate, power=0.5) model.compile(optimizer=tf.keras.optimizers.SGD( learning_rate=learning_rate_fn), loss='sparse_categorical_crossentropy', metrics=['accuracy']) model.fit(data, labels, epochs=5) ``` The learning rate schedule is also serializable and deserializable using `tf.keras.optimizers.schedules.serialize` and `tf.keras.optimizers.schedules.deserialize`. Returns: A 1-arg callable learning rate schedule that takes the current optimizer step and outputs the decayed learning rate, a scalar `Tensor` of the same type as `initial_learning_rate`. -C6??FNc2t||_||_||_||_||_||_dS)aApplies a polynomial decay to the learning rate. Args: initial_learning_rate: A scalar `float32` or `float64` `Tensor` or a Python number. The initial learning rate. decay_steps: A scalar `int32` or `int64` `Tensor` or a Python number. Must be positive. See the decay computation above. end_learning_rate: A scalar `float32` or `float64` `Tensor` or a Python number. The minimal end learning rate. power: A scalar `float32` or `float64` `Tensor` or a Python number. The power of the polynomial. Defaults to `1.0`. cycle: A boolean, whether it should cycle beyond decay_steps. name: String. Optional name of the operation. Defaults to 'PolynomialDecay'. N)r#r$r%r&end_learning_ratepowercycler))rr%r&rerfrgr)r+rrr$s  zPolynomialDecay.__init__c Cst|jpdj}tj|jdd}|j}t|j|}t|j|}t||}t|j |}|j rKt t |ddtj ||j } t|| }nt||}t||} tjt||td| |||dWdS1suwYdS)Nrar%r,rrcr?)r-r.r)r/r%r0r1rerfr&rgwhereequalmathceilr3minimumdivideaddr4) rrr)r%r0rerfr5decay_steps_recomp multiplierr6rrrrs:   $zPolynomialDecay.__call__cC|j|j|j|j|j|jdS)Nr%r&rerfrgr)rrrrrrrzPolynomialDecay.get_config)rbrcFNr<rrr+rra=sJ!%raz+keras.optimizers.schedules.InverseTimeDecaycr) InverseTimeDecayaTA LearningRateSchedule that uses an inverse time decay schedule. When training a model, it is often useful to lower the learning rate as the training progresses. This schedule applies the inverse decay function to an optimizer step, given a provided initial learning rate. It requires a `step` value to compute the decayed learning rate. You can just pass a TensorFlow variable that you increment at each training step. The schedule is a 1-arg callable that produces a decayed learning rate when passed the current optimizer step. This can be useful for changing the learning rate value across different invocations of optimizer functions. It is computed as: ```python def decayed_learning_rate(step): return initial_learning_rate / (1 + decay_rate * step / decay_step) ``` or, if `staircase` is `True`, as: ```python def decayed_learning_rate(step): return initial_learning_rate / (1 + decay_rate * floor(step / decay_step)) ``` You can pass this schedule directly into a `tf.keras.optimizers.Optimizer` as the learning rate. Example: Fit a Keras model when decaying 1/t with a rate of 0.5: ```python ... initial_learning_rate = 0.1 decay_steps = 1.0 decay_rate = 0.5 learning_rate_fn = keras.optimizers.schedules.InverseTimeDecay( initial_learning_rate, decay_steps, decay_rate) model.compile(optimizer=tf.keras.optimizers.SGD( learning_rate=learning_rate_fn), loss='sparse_categorical_crossentropy', metrics=['accuracy']) model.fit(data, labels, epochs=5) ``` Returns: A 1-arg callable learning rate schedule that takes the current optimizer step and outputs the decayed learning rate, a scalar `Tensor` of the same type as `initial_learning_rate`. FNcr!)aApplies inverse time decay to the initial learning rate. Args: initial_learning_rate: A scalar `float32` or `float64` `Tensor` or a Python number. The initial learning rate. decay_steps: How often to apply decay. decay_rate: A Python number. The decay rate. staircase: Whether to apply decay in a discrete staircase, as opposed to continuous, fashion. name: String. Optional name of the operation. Defaults to 'InverseTimeDecay'. Nr"r*r+rrr$ s  zInverseTimeDecay.__init__c Cst|jpdN}tj|jdd}|j}t|j|}t|j|}t||}||}|j r4t |}tt d|} t | t ||} tj|| |dWdS1sYwYdS)Nrtr%r,r?)r-r.r)r/r%r0r1r&r'r(r2constantrnr3rm) rrr)r%r0r&r'r5r6constdenomrrrr's  $zInverseTimeDecay.__call__cCr7r8r9rrrrr8r:zInverseTimeDecay.get_configr;r<rrr+rrts8rtz&keras.optimizers.schedules.CosineDecayzkeras.experimental.CosineDecaycsJeZdZdZ    dfdd ZddZd d Zd d Zd dZZ S) CosineDecaya A LearningRateSchedule that uses a cosine decay with optional warmup. See [Loshchilov & Hutter, ICLR2016](https://arxiv.org/abs/1608.03983), SGDR: Stochastic Gradient Descent with Warm Restarts. For the idea of a linear warmup of our learning rate, see [Goyal et al.](https://arxiv.org/pdf/1706.02677.pdf). When we begin training a model, we often want an initial increase in our learning rate followed by a decay. If `warmup_target` is an int, this schedule applies a linear increase per optimizer step to our learning rate from `initial_learning_rate` to `warmup_target` for a duration of `warmup_steps`. Afterwards, it applies a cosine decay function taking our learning rate from `warmup_target` to `alpha` for a duration of `decay_steps`. If `warmup_target` is None we skip warmup and our decay will take our learning rate from `initial_learning_rate` to `alpha`. It requires a `step` value to compute the learning rate. You can just pass a TensorFlow variable that you increment at each training step. The schedule is a 1-arg callable that produces a warmup followed by a decayed learning rate when passed the current optimizer step. This can be useful for changing the learning rate value across different invocations of optimizer functions. Our warmup is computed as: ```python def warmup_learning_rate(step): completed_fraction = step / warmup_steps total_delta = target_warmup - initial_learning_rate return completed_fraction * total_delta ``` And our decay is computed as: ```python if warmup_target is None: initial_decay_lr = initial_learning_rate else: initial_decay_lr = warmup_target def decayed_learning_rate(step): step = min(step, decay_steps) cosine_decay = 0.5 * (1 + cos(pi * step / decay_steps)) decayed = (1 - alpha) * cosine_decay + alpha return initial_decay_lr * decayed ``` Example usage without warmup: ```python decay_steps = 1000 initial_learning_rate = 0.1 lr_decayed_fn = tf.keras.optimizers.schedules.CosineDecay( initial_learning_rate, decay_steps) ``` Example usage with warmup: ```python decay_steps = 1000 initial_learning_rate = 0 warmup_steps = 1000 target_learning_rate = 0.1 lr_warmup_decayed_fn = tf.keras.optimizers.schedules.CosineDecay( initial_learning_rate, decay_steps, warmup_target=target_learning_rate, warmup_steps=warmup_steps ) ``` You can pass this schedule directly into a `tf.keras.optimizers.Optimizer` as the learning rate. The learning rate schedule is also serializable and deserializable using `tf.keras.optimizers.schedules.serialize` and `tf.keras.optimizers.schedules.deserialize`. Returns: A 1-arg callable learning rate schedule that takes the current optimizer step and outputs the decayed learning rate, a scalar `Tensor` of the same type as `initial_learning_rate`. Nrcs2t||_||_||_||_||_||_dS)aiApplies cosine decay to the learning rate. Args: initial_learning_rate: A scalar `float32` or `float64` `Tensor` or a Python int. The initial learning rate. decay_steps: A scalar `int32` or `int64` `Tensor` or a Python int. Number of steps to decay over. alpha: A scalar `float32` or `float64` `Tensor` or a Python int. Minimum learning rate value for decay as a fraction of `initial_learning_rate`. name: String. Optional name of the operation. Defaults to 'CosineDecay'. warmup_target: None or a scalar `float32` or `float64` `Tensor` or a Python int. The target learning rate for our warmup phase. Will cast to the `initial_learning_rate` datatype. Setting to None will skip warmup and begins decay phase from `initial_learning_rate`. Otherwise scheduler will warmup from `initial_learning_rate` to `warmup_target`. warmup_steps: A scalar `int32` or `int64` `Tensor` or a Python int. Number of steps to warmup over. N)r#r$r%r&alphar) warmup_steps warmup_target)rr%r&rzr)r|r{r+rrr$s  zCosineDecay.__init__c Cst|jpd/||}tjtj|d}ddt||}d|j||j}t||WdS1s:wYdS)Nrxr0?rcr?) r-r.r)rurjpicosrzr3) rrr& decay_from_lrr0completed_fractiontf_picosine_decayeddecayedrrr_decay_functions $zCosineDecay._decay_functioncCsPt|jpd||}||}|||WdS1s!wYdS)Nrx)r-r.r))rrr{r|r%rtotal_step_deltarrr_warmup_functions  $zCosineDecay._warmup_functionc stjpdjtjjddjtjt|jdur;t  WdStjtj t t kfddfddWdS1suwYdS)Nrxr%r,csSrL)rr)r5r%rr{r|rrrIs z&CosineDecay.__call__..csSrL)rr)r&r0r5rr{r|rrrIs ) r-r.r)r/r%r0r1r&r|rlrr{condr r)r&r0r5r%rr{r|rrs6   $zCosineDecay.__call__cCrq)Nr%r&rzr)r|r{rrrrrrrszCosineDecay.get_config)ryNNr) r rrrr$rrrrr=rrr+rrxBsU')rxz.keras.optimizers.schedules.CosineDecayRestartsz&keras.experimental.CosineDecayRestartscr`) CosineDecayRestartsa A LearningRateSchedule that uses a cosine decay schedule with restarts. See [Loshchilov & Hutter, ICLR2016](https://arxiv.org/abs/1608.03983), SGDR: Stochastic Gradient Descent with Warm Restarts. When training a model, it is often useful to lower the learning rate as the training progresses. This schedule applies a cosine decay function with restarts to an optimizer step, given a provided initial learning rate. It requires a `step` value to compute the decayed learning rate. You can just pass a TensorFlow variable that you increment at each training step. The schedule is a 1-arg callable that produces a decayed learning rate when passed the current optimizer step. This can be useful for changing the learning rate value across different invocations of optimizer functions. The learning rate multiplier first decays from 1 to `alpha` for `first_decay_steps` steps. Then, a warm restart is performed. Each new warm restart runs for `t_mul` times more steps and with `m_mul` times initial learning rate as the new learning rate. Example usage: ```python first_decay_steps = 1000 lr_decayed_fn = ( tf.keras.optimizers.schedules.CosineDecayRestarts( initial_learning_rate, first_decay_steps)) ``` You can pass this schedule directly into a `tf.keras.optimizers.Optimizer` as the learning rate. The learning rate schedule is also serializable and deserializable using `tf.keras.optimizers.schedules.serialize` and `tf.keras.optimizers.schedules.deserialize`. Returns: A 1-arg callable learning rate schedule that takes the current optimizer step and outputs the decayed learning rate, a scalar `Tensor` of the same type as `initial_learning_rate`. @rcryNcrd)a~Applies cosine decay with restarts to the learning rate. Args: initial_learning_rate: A scalar `float32` or `float64` Tensor or a Python number. The initial learning rate. first_decay_steps: A scalar `int32` or `int64` `Tensor` or a Python number. Number of steps to decay over. t_mul: A scalar `float32` or `float64` `Tensor` or a Python number. Used to derive the number of iterations in the i-th period. m_mul: A scalar `float32` or `float64` `Tensor` or a Python number. Used to derive the initial learning rate of the i-th period. alpha: A scalar `float32` or `float64` Tensor or a Python number. Minimum learning rate value as a fraction of the initial_learning_rate. name: String. Optional name of the operation. Defaults to 'SGDRDecay'. N)r#r$r%first_decay_steps_t_mul_m_mulrzr))rr%rt_mulm_mulrzr)r+rrr$/s  zCosineDecayRestarts.__init__c s t|jpd~}tj|jdd}|j}t|j|}t|j|}t|j |t|j |}t||}||dfdd t t dfdd fd d \} || } d | dt tjtj|d } d || |} tj|| |dWdS1swYdS)N SGDRDecayr%r,Fcst|r-ttjd|dtj}d|d}|||}||fSt|}||8}||fS)zHelper for `cond` operation.rc)r-r2rjlog)r geometric i_restartsum_r)rrr compute_step_s  z2CosineDecayRestarts.__call__..compute_steprcc ddS)NFrrrrrrrrIt z.CosineDecayRestarts.__call__..cr)NTrrrrrrrIurr~r}r?F)r-r.r)r/r%r0r1rrzrrrrirrurjrr3) rrr)r%r0rrzrr5rm_facrrr)rrrrrQs>     $zCosineDecayRestarts.__call__cCrq)N)r%rrrrzr))r%rrrrzr)rrrrrrszCosineDecayRestarts.get_config)rrcryNr<rrr+rrs,"6rcr`) LinearCosineDecayasA LearningRateSchedule that uses a linear cosine decay schedule. See [Bello et al., ICML2017] Neural Optimizer Search with RL. https://arxiv.org/abs/1709.07417 For the idea of warm starts here controlled by `num_periods`, see [Loshchilov & Hutter, ICLR2016] SGDR: Stochastic Gradient Descent with Warm Restarts. https://arxiv.org/abs/1608.03983 Note that linear cosine decay is more aggressive than cosine decay and larger initial learning rates can typically be used. When training a model, it is often recommended to lower the learning rate as the training progresses. This schedule applies a linear cosine decay function to an optimizer step, given a provided initial learning rate. It requires a `step` value to compute the decayed learning rate. You can just pass a TensorFlow variable that you increment at each training step. The schedule is a 1-arg callable that produces a decayed learning rate when passed the current optimizer step. This can be useful for changing the learning rate value across different invocations of optimizer functions. It is computed as: ```python def decayed_learning_rate(step): step = min(step, decay_steps) linear_decay = (decay_steps - step) / decay_steps cosine_decay = 0.5 * ( 1 + cos(pi * 2 * num_periods * step / decay_steps)) decayed = (alpha + linear_decay) * cosine_decay + beta return initial_learning_rate * decayed ``` Example usage: ```python decay_steps = 1000 lr_decayed_fn = ( tf.keras.experimental.LinearCosineDecay( initial_learning_rate, decay_steps)) ``` You can pass this schedule directly into a `tf.keras.optimizers.Optimizer` as the learning rate. The learning rate schedule is also serializable and deserializable using `tf.keras.optimizers.schedules.serialize` and `tf.keras.optimizers.schedules.deserialize`. Returns: A 1-arg callable learning rate schedule that takes the current optimizer step and outputs the decayed learning rate, a scalar `Tensor` of the same type as `initial_learning_rate`. r~ryMbP?Ncrd)ajApplies linear cosine decay to the learning rate. Args: initial_learning_rate: A scalar `float32` or `float64` Tensor or a Python number. The initial learning rate. decay_steps: A scalar `int32` or `int64` `Tensor` or a Python number. Number of steps to decay over. num_periods: Number of periods in the cosine part of the decay. See computation above. alpha: See computation above. beta: See computation above. name: String. Optional name of the operation. Defaults to 'LinearCosineDecay'. N)r#r$r%r& num_periodsrzbetar))rr%r&rrzrr)r+rrr$s  zLinearCosineDecay.__init__c Cst|jpdl}tj|jdd}|j}t|j|}t|j|}t|j |}t|j |}t||} t | |} || |} | |} d|| } ddt tj tj|d| } || | |}tj|||dWdS1swwYdS)Nrr%r,rr~rcr})r-r.r)r/r%r0r1r&rrzrrlrrurjrr3)rrr)r%r0r&rrzrr5linear_decayedrfractionrlinear_cosine_decayedrrrrs4    $zLinearCosineDecay.__call__cCrq)Nr%r&rrzrr)rrrrrrrszLinearCosineDecay.get_config)r~ryrNr<rrr+rrs8 rcs@eZdZdZ       dfdd Zd d Zd d ZZS)NoisyLinearCosineDecaya A LearningRateSchedule that uses a noisy linear cosine decay schedule. See [Bello et al., ICML2017] Neural Optimizer Search with RL. https://arxiv.org/abs/1709.07417 For the idea of warm starts here controlled by `num_periods`, see [Loshchilov & Hutter, ICLR2016] SGDR: Stochastic Gradient Descent with Warm Restarts. https://arxiv.org/abs/1608.03983 Note that linear cosine decay is more aggressive than cosine decay and larger initial learning rates can typically be used. When training a model, it is often recommended to lower the learning rate as the training progresses. This schedule applies a noisy linear cosine decay function to an optimizer step, given a provided initial learning rate. It requires a `step` value to compute the decayed learning rate. You can just pass a TensorFlow variable that you increment at each training step. The schedule is a 1-arg callable that produces a decayed learning rate when passed the current optimizer step. This can be useful for changing the learning rate value across different invocations of optimizer functions. It is computed as: ```python def decayed_learning_rate(step): step = min(step, decay_steps) linear_decay = (decay_steps - step) / decay_steps) cosine_decay = 0.5 * ( 1 + cos(pi * 2 * num_periods * step / decay_steps)) decayed = (alpha + linear_decay + eps_t) * cosine_decay + beta return initial_learning_rate * decayed ``` where eps_t is 0-centered gaussian noise with variance initial_variance / (1 + global_step) ** variance_decay Example usage: ```python decay_steps = 1000 lr_decayed_fn = ( tf.keras.experimental.NoisyLinearCosineDecay( initial_learning_rate, decay_steps)) ``` You can pass this schedule directly into a `tf.keras.optimizers.Optimizer` as the learning rate. The learning rate schedule is also serializable and deserializable using `tf.keras.optimizers.schedules.serialize` and `tf.keras.optimizers.schedules.deserialize`. Returns: A 1-arg callable learning rate schedule that takes the current optimizer step and outputs the decayed learning rate, a scalar `Tensor` of the same type as `initial_learning_rate`. rc皙?r~ryrNc sPt||_||_||_||_||_||_||_||_ | |_ t ||_ dS)auApplies noisy linear cosine decay to the learning rate. Args: initial_learning_rate: A scalar `float32` or `float64` Tensor or a Python number. The initial learning rate. decay_steps: A scalar `int32` or `int64` `Tensor` or a Python number. Number of steps to decay over. initial_variance: initial variance for the noise. See computation above. variance_decay: decay for the noise's variance. See computation above. num_periods: Number of periods in the cosine part of the decay. See computation above. alpha: See computation above. beta: See computation above. seed: Integer, optional random seed to enable deterministic behavior. name: String. Optional name of the operation. Defaults to 'NoisyLinearCosineDecay'. N)r#r$r%r&initial_variancevariance_decayrrzrseedr)rRandomGenerator_random_generator) rr%r&rrrrzrrr)r+rrr$Fs zNoisyLinearCosineDecay.__init__c CsLt|jpd}tj|jdd}|j}t|j|}t|j|}t|j |}t|j |}t|j |} t|j |} t||} t | |} || |} |td| |} t| }| |jj| j|d}| |}d||}ddttjtj|d|}| ||| }tj|||dWdS1swYdS) Nrr%r,rc)stddevrr~r})r-r.r)r/r%r0r1r&rrrrzrrlr4sqrtr random_normalshaperrurjrr3)rrr)r%r0r&rrrrzrr5rvariancestdnoisy_linear_decayedrrrnoisy_linear_cosine_decayedrrrrqsL     $zNoisyLinearCosineDecay.__call__c Cs*|j|j|j|j|j|j|j|j|jd S)N r%r&rrrrzrrr)rrrrrrsz!NoisyLinearCosineDecay.get_config)rcrr~ryrNNr<rrr+rrs:+(rz$keras.optimizers.schedules.serializeFcCs|rt|St|S)aDSerializes a `LearningRateSchedule` into a JSON-compatible dict. Args: learning_rate_schedule: The `LearningRateSchedule` object to serialize. Returns: A JSON-serializable dict representing the object's config. Example: >>> lr_schedule = tf.keras.optimizers.schedules.ExponentialDecay( ... 0.1, decay_steps=100000, decay_rate=0.96, staircase=True) >>> tf.keras.optimizers.schedules.serialize(lr_schedule) {'module': 'keras.optimizers.schedules', 'class_name': 'ExponentialDecay', 'config': {...}, 'registered_name': None} )legacy_serializationserialize_keras_objectr)learning_rate_scheduleuse_legacy_formatrrr serializes  rz&keras.optimizers.schedules.deserializecCs,|r tj|t|ddStj|t|ddS)aInstantiates a `LearningRateSchedule` object from a serialized form. Args: config: The serialized form of the `LearningRateSchedule`. Dictionary of the form {'class_name': str, 'config': dict}. custom_objects: A dictionary mapping class names (or function names) of custom (non-Keras) objects to class/functions. Returns: A `LearningRateSchedule` object. Example: ```python # Configuration for PolynomialDecay config = { 'class_name': 'PolynomialDecay', 'config': {'cycle': False, 'decay_steps': 10000, 'end_learning_rate': 0.01, 'initial_learning_rate': 0.1, 'name': None, 'power': 0.5}} lr_schedule = tf.keras.optimizers.schedules.deserialize(config) ``` decay)module_objectscustom_objectsprintable_module_name)rdeserialize_keras_objectglobalsr)rrrrrr deserializesrr)NF)rrrjtensorflow.compat.v2compatv2r- keras.srcrkeras.src.savingrkeras.src.saving.legacyrr tensorflow.python.util.tf_exportrrr r>rartrxrrrrrrrrrsJ    Fkkk>|