hh"dZddlmcmZddlmZddlmZddl m Z ee dddg Gd d ejZ e jjd eje _y) zAdamW optimizer implementation.N) optimizer)register_keras_serializable) keras_exportzkeras.optimizers.AdamWz#keras.optimizers.experimental.AdamWz+keras.dtensor.experimental.optimizers.AdamW)v1cZeZdZdZ dfd ZfdZdZfdZxZS)AdamWa Optimizer that implements the AdamW algorithm. AdamW optimization is a stochastic gradient descent method that is based on adaptive estimation of first-order and second-order moments with an added method to decay weights per the techniques discussed in the paper, 'Decoupled Weight Decay Regularization' by [Loshchilov, Hutter et al., 2019](https://arxiv.org/abs/1711.05101). According to [Kingma et al., 2014](http://arxiv.org/abs/1412.6980), the underying Adam method is "*computationally efficient, has little memory requirement, invariant to diagonal rescaling of gradients, and is well suited for problems that are large in terms of data/parameters*". Args: learning_rate: A `tf.Tensor`, floating point value, a schedule that is a `tf.keras.optimizers.schedules.LearningRateSchedule`, or a callable that takes no arguments and returns the actual value to use. The learning rate. Defaults to 0.001. beta_1: A float value or a constant float tensor, or a callable that takes no arguments and returns the actual value to use. The exponential decay rate for the 1st moment estimates. Defaults to 0.9. beta_2: A float value or a constant float tensor, or a callable that takes no arguments and returns the actual value to use. The exponential decay rate for the 2nd moment estimates. Defaults to 0.999. epsilon: A small constant for numerical stability. This epsilon is "epsilon hat" in the Kingma and Ba paper (in the formula just before Section 2.1), not the epsilon in Algorithm 1 of the paper. Defaults to 1e-7. amsgrad: Boolean. Whether to apply AMSGrad variant of this algorithm from the paper "On the Convergence of Adam and beyond". Defaults to `False`. {{base_optimizer_keyword_args}} Reference: - [Loshchilov et al., 2019](https://arxiv.org/abs/1711.05101) - [Kingma et al., 2014](http://arxiv.org/abs/1412.6980) for `adam` - [Reddi et al., 2018]( https://openreview.net/pdf?id=ryQu7f-RZ) for `amsgrad`. Notes: The sparse implementation of this algorithm (used when the gradient is an IndexedSlices object, typically because of `tf.gather` or an embedding lookup in the forward pass) does apply momentum to variable slices even if they were not used in the forward pass (meaning they have a gradient equal to zero). Momentum decay (beta1) is also applied to the entire momentum accumulator. This means that the sparse behavior is equivalent to the dense behavior (in contrast to some momentum implementations which ignore momentum unless a variable slice was actually used). c t|d|||| | | | | d||j||_||_||_||_||_||_|j tdy)N)nameclipnorm clipvalueglobal_clipnormuse_ema ema_momentumema_overwrite_frequency jit_compilezLMissing value of `weight_decay` which is required and must be a float value.) super__init___build_learning_rate_learning_rate weight_decaybeta_1beta_2epsilonamsgrad ValueError)self learning_raterrrrrr r r rrrrr kwargs __class__s [/var/www/html/dev/engine/venv/lib/python3.12/site-packages/tf_keras/src/optimizers/adamw.pyrzAdamW.__init__Zs$  +%$;#  #77 F(       $*  %ct||t|dr |jryd|_g|_g|_|D]Z}|jj |j|d|j j |j|d\|jr;g|_ |D].}|jj |j|d0yy)aInitialize optimizer variables. AdamW optimizer has 3 types of variables: momentums, velocities and velocity_hat (only set when amsgrad is applied), Args: var_list: list of model variables to build AdamW variables on. _builtNTm)model_variable variable_namevvhat) rbuildhasattrr$ _momentums _velocitiesappendadd_variable_from_referencer_velocity_hats)rvar_listvarr s r!r*z AdamW.builds  h 4 "t{{   C OO " "00#&c1     # #00#&c1   <<"$D  ##**44'*&5  r"cVtj|j|j}tj|jdz|j}tj tj|j |j|}tj tj|j|j|}|j|}|j|j|}|j|j|} |tjd|z zd|z z } t|tjrp|j| d|j z z|j!tj|j"d|j z z|j$| j| d|jz z| j!tjtj&|j"d|jz z|j$|j(rC|j*|j|} | j-tj.| | | } |j1|| ztj| |j2zz y|j||z d|j z z| jtj&|| z d|jz z|j(rC|j*|j|} | j-tj.| | | } |j1|| ztj| |j2zz y)z=Update step given gradient and the associated model variable.N)tfcastrdtype iterationspowrr_var_keyr, _index_dictr-sqrt isinstance IndexedSlices assign_add scatter_addvaluesindicessquarerr0assignmaximum assign_subr) rgradientvariablelr local_step beta_1_power beta_2_powervar_keyr%r(alphav_hats r! update_stepzAdamW.update_steps WWT'' 8WWT__q0(..A vvbggdkk8>>BJO vvbggdkk8>>BJO --) OOD,,W5 6   T--g6 7RWWQ-..!l2BC h 0 0 1 LL!q4;;/ 0 MM  OOq4;;79I9I  LL!q4;;/ 0 MM  IIhoo.!dkk/B$$  ||++D,<,FNNNFgGz?NTr) __name__ __module__ __qualname____doc__rr*rPrR __classcell__)r s@r!rrsR5r $ (T!F)KVr"rz{{base_optimizer_keyword_args}})rYtensorflow.compat.v2compatv2r5tf_keras.src.optimizersr'tf_keras.src.saving.object_registrationr tensorflow.python.util.tf_exportr Optimizerrreplacebase_optimizer_keyword_argsrr"r!rds&"!-O: )1  I   D %%%y'L'L r"