BVh!dZddlmZddlmZddlmZddlm Z ddl m Z ddl m Z ddlmZed g Gd d e j Zy )zAOptimizer that implements cross-shard gradient reduction for TPU.)ops)losses) tf_logging) tpu_function)tpu_ops) optimizer) tf_exportztpu.CrossShardOptimizer)v1czeZdZdZej j ddffd ZdZd dZ d dZ dZ dZ d Z xZS) CrossShardOptimizerz7An optimizer that averages gradients across TPU shards.Nc<tjjtjjf}||vrt d|d|t |t jstd|dtt|/d|||_ ||_ ||_y)aConstruct a new cross-shard optimizer. Args: opt: An existing `Optimizer` to encapsulate. reduction: The reduction to apply to the shard losses. name: Optional name prefix for the operations created when applying gradients. Defaults to "CrossShardOptimizer". group_assignment: Optional 2d int32 lists with shape [num_groups, num_replicas_per_group] which describles how to apply optimizer to subgroups. Raises: ValueError: If reduction is not a valid cross-shard reduction. z&Argument `reduction` should be one of z . Received: z]CrossShardOptimizer only works with tf.training.Optimizer and not Keras Optimizer. Received: z. If you are using TPUStrategy, Keras Optimizer will sum gradients across replicas.If you want to average your gradients, rescale your loss with: `loss /= global_batch_size`FN)r ReductionSUMMEAN ValueError isinstancer Optimizer TypeErrorsuperr __init___opt _reduction_group_assignment)selfopt reductionnamegroup_assignmentaccepted_reductions __class__s S/home/dcms/DCMS/lib/python3.12/site-packages/tensorflow/python/tpu/tpu_optimizer.pyrzCrossShardOptimizer.__init__ s&"++//1A1A1F1FG++  23F2GH k # $$ c9.. /  ((+u-( ( )) t-eT:DIDO-Dc|syt|trtd|Dstd|t }|D]}|D]}|j |t t ||k7rtd|d||Dcgc] }t|c}tfdDrdStd|cc}w) aKVerify group_assignment and get the subgroup size". Args: group_assignment: list of group ids for applying the optimizer to subgroups. num_shards: The number of TPU shards. Returns: The size of one subgroup in group_assignment. Raises: ValueError: If group_assignment is invalid. Nc3<K|]}t|tywN)rlist).0is r! zDCrossShardOptimizer._verify_and_get_subgroup_size..Ws> 1d#>sz?Argument `group_assignment` must be a list of lists. Received: z;Argument `group_assignment` must be a permutation of range(z ). Received: c3.K|] }d|k(yw)rN)r'sizesubgroup_size_lists r!r)zDCrossShardOptimizer._verify_and_get_subgroup_size..gs HT a D ( HsrzIThe size of each subgroup in `group_assignment` must be equal. Received: )rr&allrsetaddrangelen)rr num_shards replica_idsgr(groupr-s @r!_verify_and_get_subgroup_sizez1CrossShardOptimizer._verify_and_get_subgroup_sizeFs   ' . >-=> >  '( * ++%K ! 5 ,  ,m,<+= ? @@3CC#e*C H5G HH  "" ..>-?A BB Ds Cc vtjj}|tjdd}|j |j |}|dkDrC|jtjjk(r|j rd|z }nd|z }||z}|jj|fd|i|S)a1Compute gradients of "loss" for the variables in "var_list". This simply wraps `compute_gradients()` from the real optimizer. The gradients will be aggregated in `apply_gradients()` so that user can modify the gradients like clipping with per replica global norm if needed. The global norm with aggregated gradients can be bad as one replica's huge gradients can hurt the gradients from other replicas. When the CrossShardOptimizer is constructed with `reduction == losses.Reduction.MEAN` (default), this function scales the loss by `1.0 / num_shards` before computing the gradients. Assuming the optimizer uses the default implementation of `compute_gradients()`, the gradients of the scaled loss are scaled by `1.0 / num_shards` compared to the gradients of the original loss. This scaling factor is important because `apply_gradients()` sums gradients across shards, rather than averaging them. However, the scaling factor must be taken into account when clipping the norm of the gradients or performing other postprocessing. Args: loss: A Tensor containing the value to minimize. var_list: Optional list or tuple of `tf.Variable` to update to minimize `loss`. Defaults to the list of variables collected in the graph under the key `GraphKey.TRAINABLE_VARIABLES`. **kwargs: Keyword arguments for compute_gradients(). Returns: A list of (gradient, variable) pairs. Raises: ValueError: If not within a tpu_shard_context or group_assignment is invalid. zjCrossShardOptimizer should be used within a tpu_shard_context, but got unset number_of_shards. Assuming 1.g?var_list) rget_tpu_contextnumber_of_shardsloggingwarningr7rrrrrrcompute_gradients)rlossr:kwargsr3 subgroup_sizescales r!r?z%CrossShardOptimizer.compute_gradientsmsB--/@@J oo 45j66t7M7M7ACMA~$//V-=-=-B-BB  m#j  emd &499 & &t Ih I& IIr"c6g}|D]j\}}||j||ftj|5|jtj||j |fdddl|j j|||S#1swYxYw)aApply gradients to variables. Calls tpu_ops.cross_replica_sum() to sum gradient contributions across replicas, and then applies the real optimizer. Args: grads_and_vars: List of (gradient, variable) pairs as returned by compute_gradients(). global_step: Optional Variable to increment by one after the variables have been updated. name: Optional name for the returned operation. Default to the name passed to the Optimizer constructor. Returns: An `Operation` that applies the gradients. If `global_step` was not None, that operation also increments `global_step`. Raises: ValueError: If the grads_and_vars is malformed. N)appendr colocate_withrcross_replica_sumrrapply_gradients)rgrads_and_vars global_steprsummed_grads_and_varsgradvars r!rHz#CrossShardOptimizer.apply_gradientss*%3 s $$dC[1   t $ 3  & &(A(AD**),-0(2 3 3 3 3 99 $ $%:K NN 3 3s 2BB c:|jj|i|S)a=Return a slot named "name" created for "var" by the Optimizer. This simply wraps the get_slot() from the actual optimizer. Args: *args: Arguments for get_slot(). **kwargs: Keyword arguments for get_slot(). Returns: The `Variable` for the slot if it was created, `None` otherwise. )rget_slotrargsrAs r!rOzCrossShardOptimizer.get_slots  499  t .v ..r"c:|jj|i|S)aReturn a list of the names of slots created by the `Optimizer`. This simply wraps the get_slot_names() from the actual optimizer. Args: *args: Arguments for get_slot(). **kwargs: Keyword arguments for get_slot(). Returns: A list of strings. )rget_slot_namesrPs r!rSz"CrossShardOptimizer.get_slot_namess  $499 # #T 4V 44r"c6|jjS)z7Forwarding the variables from the underlying optimizer.)r variables)rs r!rUzCrossShardOptimizer.variabless 99    r"r%)NN)__name__ __module__ __qualname____doc__rrrrr7r?rHrOrSrU __classcell__)r s@r!r r sF? ))..) $ $.L%BN2JhO> / 5!r"r N)rYtensorflow.python.frameworkrtensorflow.python.ops.lossesrtensorflow.python.platformrr=tensorflow.python.tpurtensorflow.python.tpu.opsrtensorflow.python.trainingr tensorflow.python.util.tf_exportr rr r+r"r!rbsO H,/<.-06 ()*A!)--A!+A!r"