AVhbdZddlmZddlmZGddeZGddeZGdd eZ y ) z@The step function abstraction represents a single training step.)backprop) optimizerc2eZdZdZdZedZdZdZy)Stepz;Interface for performing each step of a training algorithm.c||_yN _distribution)self distributions T/home/dcms/DCMS/lib/python3.12/site-packages/tensorflow/python/distribute/step_fn.py__init__z Step.__init__s %Dc|jSrr r s r r zStep.distributions   rcgSrrs r initializezStep.initializes Irctd)z,Perform one step of this training algorithm.z"must be implemented in descendants)NotImplementedErrorrs r __call__z Step.__call__"s B CCrN) __name__ __module__ __qualname____doc__rpropertyr rrrrr rrs)C&  Drrc(eZdZdZfdZdZxZS)StandardInputStepzStep with a standard implementation of input handling. Args: dataset_fn: a function that returns a tf.data Dataset that produces the input for the model. c`tt| ||jfd|_y)NcSrr)_ dataset_fns r z,StandardInputStep.__init__..3s :<r)superrrmake_input_fn_iterator _iterator)r r"r __class__s ` r rzStandardInputStep.__init__1s& T+L9!889OPDNrc.|jjSr)r& initializerrs r rzStandardInputStep.initialize5s >> % %%r)rrrrrr __classcell__r's@r rr)sQ&rrc,eZdZdZ dfd ZdZxZS)StandardSingleLossStepaA step function that implements a training step for a feed forward network. An instance of this class is intended to be used as a callable: ```python ... step = step_fn.StandardSingleLossStep( dataset, loss_fn, optimizer, distribution) # Run a single training step on a given DistributionStrategy: step(distribution) ... ``` Args: dataset_fn: a function that returns a tf.data Dataset that produces the input for the model. loss_fn: a function that takes a context and inputs as arguments. It returns the loss for those inputs. `context` is an instance of `values.MultiStepContext` that will be passed when `loss_fn` is run. `context` can be used to specify the outputs to be returned from `loss_fn`, among other things. optimizer: an optimizer that implements an update rule. distribution: a `DistributionStrategy` object. cXtt| ||||_||_||_yr)r$r-r_loss_fn _optimizer_iterations_per_step)r r"loss_fnrr iterations_per_stepr's r rzStandardSingleLossStep.__init__Ts+ $0\JDMDO 3Drcjj5fd}jjj |j j }|jcdddS#1swYyxYw)Nc tjj}tj|}j j j|||f}jjj |S)z-Function to run one iteration with one input.)args) r implicit_gradr/ optimizer_libget_filtered_grad_fnr extendedcall_for_each_replicar0_distributed_apply)ctxinputs gradients_fngrads_and_varsr s r step_fnz0StandardSingleLossStep.__call__..step_fn]sv--dmm< $99,G **33II V}J. 11   ~/ /r)r scoper r:"experimental_run_steps_on_iteratorr&r1run_op)r rAr=s` r rzStandardSingleLossStep.__call__[se   ! ! # /    & & I I 4>>4#<#< >c ZZ%s A A22A;))rrrrrrr*r+s@r r-r-9s6$%4rr-N) rtensorflow.python.eagerrtensorflow.python.trainingrr8objectrrr-rrr rIs8G,AD6D( & & 5.5r