SendTPUEmbeddingGradients

public final class SendTPUEmbeddingGradients

Performs gradient updates of embedding tables.

Public Methods

static SendTPUEmbeddingGradients
create(Scope scope, Iterable<Operand<Float>> inputs, Iterable<Operand<Float>> learningRates, String config)
Factory method to create a class wrapping a new SendTPUEmbeddingGradients operation.

Inherited Methods

org.tensorflow.op.PrimitiveOp
final boolean
equals(Object obj)
final int
hashCode()
Operation
op()
Returns the underlying Operation
final String
toString()
boolean
equals(Object arg0)
final Class<?>
getClass()
int
hashCode()
final void
notify()
final void
notifyAll()
String
toString()
final void
wait(long arg0, int arg1)
final void
wait(long arg0)
final void
wait()

Public Methods

public static SendTPUEmbeddingGradients create (Scope scope, Iterable<Operand<Float>> inputs, Iterable<Operand<Float>> learningRates, String config)

Factory method to create a class wrapping a new SendTPUEmbeddingGradients operation.

Parameters
scope current scope
inputs A TensorList of gradients with which to update embedding tables. This argument has the same length and shapes as the return value of RecvTPUEmbeddingActivations, but contains gradients of the model's loss with respect to the embedding activations. The embedding tables are updated from these gradients via the optimizer specified in the TPU embedding configuration given to tpu.initialize_system.
learningRates A TensorList of float32 scalars, one for each dynamic learning rate tag: see the comments in //third_party/tensorflow/core/protobuf/tpu/optimization_parameters.proto. Multiple tables can share the same dynamic learning rate tag as specified in the configuration. If the learning rates for all tables are constant, this list should be empty.
config Serialized TPUEmbeddingConfiguration proto.
Returns
  • a new instance of SendTPUEmbeddingGradients