ML Community Day is November 9! Join us for updates from TensorFlow, JAX, and more Learn more


An optimizer that averages gradients across TPU shards.

Inherits From: Optimizer

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.

ValueError If reduction is not a valid cross-shard reduction.



View source

Apply gradients to variables.

Calls tpu_ops.cross_replica_sum() to sum gradient contributions across replicas, and then applies the real optimizer.

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.

An Operation that applies the gradients. If global_step was not None, that operation also increments global_step.

ValueError If the grads_and_vars is malformed.


View source

Compute 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.

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().