o ZhrA@sdZddlZddlmZmZmZddlZzddlm Z Wne e fy-ddl m Z Ynwddl mZeejjdrAejjjZnejjZGdd d ejZ    d!dedededededededeedeedededeeefddZGddde ZGdd d ZdS)"z?Functions and classes related to optimization (weight updates).N)CallableOptionalUnion)Adam)keraslearning_rate_schedulec sPeZdZdZ  ddededededeef fd d Z d d Z d dZ Z S)WarmUpa Applies a warmup schedule on a given learning rate decay schedule. Args: initial_learning_rate (`float`): The initial learning rate for the schedule after the warmup (so this will be the learning rate at the end of the warmup). decay_schedule_fn (`Callable`): The schedule function to apply after the warmup for the rest of training. warmup_steps (`int`): The number of steps for the warmup part of training. power (`float`, *optional*, defaults to 1.0): The power to use for the polynomial warmup (defaults is a linear warmup). name (`str`, *optional*): Optional name prefix for the returned tensors during the schedule. ?Ninitial_learning_ratedecay_schedule_fn warmup_stepspowernamecs,t||_||_||_||_||_dSN)super__init__r r rr r)selfr r r rr __class__K/var/www/auris/lib/python3.10/site-packages/transformers/optimization_tf.pyr8s  zWarmUp.__init__cstjpd:}ttj}tjtj}||}jtj|j tj ||kfddfdd|dWdS1sEwYdS)Nr csSrrr)warmup_learning_raterrQsz!WarmUp.__call__..csjSr)r r r)rsteprrrRsr) tfZ name_scopercastZfloat32r r mathpowrZcond)rrrZglobal_step_floatZwarmup_steps_floatZwarmup_percent_doner)rrrr__call__Gs  $zWarmUp.__call__cCs|j|j|j|j|jdS)Nr r r rrr!rrrr get_configVs zWarmUp.get_config)r N) __name__ __module__ __qualname____doc__floatrintrstrrr r# __classcell__rrrrr &s"r ?+?:0yE>r init_lrnum_train_stepsnum_warmup_steps min_lr_ratio adam_beta1 adam_beta2 adam_epsilon adam_clipnormadam_global_clipnormweight_decay_raterinclude_in_weight_decayc  Csztj|||||| d} |rt|| |d} | dkr-t| | |||||gd| d } | | fStjj| |||||d} | | fS)a Creates an optimizer with a learning rate schedule using a warmup phase followed by a linear decay. Args: init_lr (`float`): The desired learning rate at the end of the warmup phase. num_train_steps (`int`): The total number of training steps. num_warmup_steps (`int`): The number of warmup steps. min_lr_ratio (`float`, *optional*, defaults to 0): The final learning rate at the end of the linear decay will be `init_lr * min_lr_ratio`. adam_beta1 (`float`, *optional*, defaults to 0.9): The beta1 to use in Adam. adam_beta2 (`float`, *optional*, defaults to 0.999): The beta2 to use in Adam. adam_epsilon (`float`, *optional*, defaults to 1e-8): The epsilon to use in Adam. adam_clipnorm (`float`, *optional*, defaults to `None`): If not `None`, clip the gradient norm for each weight tensor to this value. adam_global_clipnorm (`float`, *optional*, defaults to `None`) If not `None`, clip gradient norm to this value. When using this argument, the norm is computed over all weight tensors, as if they were concatenated into a single vector. weight_decay_rate (`float`, *optional*, defaults to 0): The weight decay to use. power (`float`, *optional*, defaults to 1.0): The power to use for PolynomialDecay. include_in_weight_decay (`List[str]`, *optional*): List of the parameter names (or re patterns) to apply weight decay to. If none is passed, weight decay is applied to all parameters except bias and layer norm parameters. )r Z decay_stepsZend_learning_rater)r r r r,)Z LayerNormZ layer_normZbias) learning_rater9beta_1beta_2epsilonclipnormglobal_clipnormexclude_from_weight_decayr:)r;r<r=r>r?r@) schedulesZPolynomialDecayr AdamWeightDecayr optimizersr)r0r1r2r3r4r5r6r7r8r9rr:Z lr_scheduleZ optimizerrrrcreate_optimizer`sD. rEcseZdZdZ         d&d eeejfd ed ed ed edede e e de e e de ffdd Z e fddZfddZddZd'fdd ZddZd'fdd Zd'fd d! Zfd"d#Zd$d%ZZS)(rCa] Adam enables L2 weight decay and clip_by_global_norm on gradients. Just adding the square of the weights to the loss function is *not* the correct way of using L2 regularization/weight decay with Adam, since that will interact with the m and v parameters in strange ways as shown in [Decoupled Weight Decay Regularization](https://arxiv.org/abs/1711.05101). Instead we want to decay the weights in a manner that doesn't interact with the m/v parameters. This is equivalent to adding the square of the weights to the loss with plain (non-momentum) SGD. Args: learning_rate (`Union[float, LearningRateSchedule]`, *optional*, defaults to 0.001): The learning rate to use or a schedule. beta_1 (`float`, *optional*, defaults to 0.9): The beta1 parameter in Adam, which is the exponential decay rate for the 1st momentum estimates. beta_2 (`float`, *optional*, defaults to 0.999): The beta2 parameter in Adam, which is the exponential decay rate for the 2nd momentum estimates. epsilon (`float`, *optional*, defaults to 1e-07): The epsilon parameter in Adam, which is a small constant for numerical stability. amsgrad (`bool`, *optional*, defaults to `False`): Whether to apply AMSGrad variant of this algorithm or not, see [On the Convergence of Adam and Beyond](https://arxiv.org/abs/1904.09237). weight_decay_rate (`float`, *optional*, defaults to 0.0): The weight decay to apply. include_in_weight_decay (`List[str]`, *optional*): List of the parameter names (or re patterns) to apply weight decay to. If none is passed, weight decay is applied to all parameters by default (unless they are in `exclude_from_weight_decay`). exclude_from_weight_decay (`List[str]`, *optional*): List of the parameter names (or re patterns) to exclude from applying weight decay to. If a `include_in_weight_decay` is passed, the names in it will supersede this list. name (`str`, *optional*, defaults to `"AdamWeightDecay"`): Optional name for the operations created when applying gradients. kwargs (`Dict[str, Any]`, *optional*): Keyword arguments. Allowed to be {`clipnorm`, `clipvalue`, `lr`, `decay`}. `clipnorm` is clip gradients by norm; `clipvalue` is clip gradients by value, `decay` is included for backward compatibility to allow time inverse decay of learning rate. `lr` is included for backward compatibility, recommended to use `learning_rate` instead. MbP?r-r.Hz>Fr,Nr;r<r=r>amsgradr9r:rArc  s4tj|||||| fi| ||_||_||_dSr)rrr9_include_in_weight_decay_exclude_from_weight_decay) rr;r<r=r>rHr9r:rArkwargsrrrrs  zAdamWeightDecay.__init__csdti}tj||dS)z?Creates an optimizer from its config with WarmUp custom object.r )custom_objects)r r from_config)clsconfigrLrrrrMszAdamWeightDecay.from_configcs0t|||tj|jdd|||fd<dS)NZadam_weight_decay_raterr9)r_prepare_localrconstantr9)r var_device var_dtype apply_staterrrrPszAdamWeightDecay._prepare_localcCsB||j}|r|j||||j|jjfd|jdStS)Nr9)Z use_locking) _do_use_weight_decayrZ assign_subdevicedtype base_dtypeZ _use_lockingrZno_op)rvarr;rTZdo_decayrrr_decay_weights_ops z!AdamWeightDecay._decay_weights_opc s.tt|\}}tjt||fd|i|S)Nr)listziprapply_gradients)rZgrads_and_varsrrKZgradstvarsrrrr]szAdamWeightDecay.apply_gradientscCs\|dur |j|ifS|pi}|||f}|dur&|||}||||f<|dd|ifS)z1Retrieves the learning rate with the given state.Nlr_trT)Z _decayed_lr_tgetZ_fallback_apply_state)rrRrSrTZ coefficientsrrr_get_lrs  zAdamWeightDecay._get_lrcsl||j|jj|\}}||||}t|gtj||fi|WdS1s/wYdSr) rarVrWrXrZrcontrol_dependenciesr_resource_apply_dense)rgradrYrTr_rKdecayrrrrcs $z%AdamWeightDecay._resource_apply_densecsn||j|jj|\}}||||}t|gtj|||fi|WdS1s0wYdSr) rarVrWrXrZrrbr_resource_apply_sparse)rrdrYindicesrTr_rKrerrrrfs $z&AdamWeightDecay._resource_apply_sparsecst}|d|ji|S)Nr9)rr#updater9)rrOrrrr#s zAdamWeightDecay.get_configcCsb|jdkrdS|jr|jD] }t||durdSq |jr/|jD] }t||dur.dSq!dS)z0Whether to use L2 weight decay for `param_name`.rFNT)r9rIresearchrJ)r param_namerrrrrU$s   z$AdamWeightDecay._do_use_weight_decay) rFr-r.rGFr,NNrCr)r$r%r&r'rr(rBLearningRateScheduleboolrr[r*r classmethodrMrPrZr]rarcrfr#rUr+rrrrrCsP(        rCc@s@eZdZdZddZeddZeddZdd Zd d Z d S) GradientAccumulatoraR Gradient accumulation utility. When used with a distribution strategy, the accumulator should be called in a replica context. Gradients will be accumulated locally on each replica and without synchronization. Users should then call `.gradients`, scale the gradients if required, and pass the result to `apply_gradients`. cCsg|_d|_dS)zInitializes the accumulator.N) _gradients _accum_stepsr"rrrrAs zGradientAccumulator.__init__cCs<|jdurtjtjdtjddtjjtjjd|_|j S)zNumber of accumulated steps.Nr)rWFZ trainableZsynchronizationZ aggregation) rrrVariablerQZint64VariableSynchronizationON_READVariableAggregationONLY_FIRST_REPLICAvaluer"rrrrFs  zGradientAccumulator.stepcCs|jstddd|jDS)z1The accumulated gradients on the current replica.zBThe accumulator should be called first to initialize the gradientscSs g|] }|dur |n|qSr)ry.0gradientrrr Xs z1GradientAccumulator.gradients..)rq ValueErrorr"rrr gradientsSszGradientAccumulator.gradientscCs|js|j}|jdd|Dt|t|jkr)tdt|jdt|t|j|D]\}}|dur@|dur@||q/|jddS)z/Accumulates `gradients` on the current replica.cSs8g|]}|durtjt|dtjjtjjdn|qS)NFrs)rrt zeros_likerurvrwrxrzrrrr}_s z0GradientAccumulator.__call__..z Expected z gradients, but got Nr)rqrextendlenr~r\Z assign_addrr)rr_Zaccum_gradientr|rrrr Zs   zGradientAccumulator.__call__cCs>|jsdS|jd|jD]}|dur|t|qdS)z8Resets the accumulated gradients on the current replica.Nr)rqrrZassignrr)rr|rrrresetts  zGradientAccumulator.resetN) r$r%r&r'rpropertyrrr rrrrrrp6s    rp) r,r-r.r/NNr,r N)r'ritypingrrrZ tensorflowrZtf_keras.optimizers.legacyr ImportErrorModuleNotFoundErrorZ"tensorflow.keras.optimizers.legacyZmodeling_tf_utilsrhasattrrDrBrrmr r(r)r[r*rErCrprrrrsf  >    T