An op that receives embedding activations on the TPU.
The TPU system performs the embedding lookups and aggregations. The results of these aggregations are visible to the Tensorflow Graph as the outputs of a XlaRecvTPUEmbeddingActivations Op. This op returns a list containing one Tensor of activations per table specified in the model.
Public Methods
static XlaRecvTPUEmbeddingActivations | |
Iterator<Operand<Float>> |
iterator()
|
List<Output<Float>> |
outputs()
A TensorList of embedding activations containing one Tensor per
embedding table in the model.
|
Inherited Methods
Public Methods
public static XlaRecvTPUEmbeddingActivations create (Scope scope, Operand<?> deduplicationData, Long numTables, String config)
Factory method to create a class wrapping a new XlaRecvTPUEmbeddingActivations operation.
Parameters
scope | current scope |
---|---|
deduplicationData | A Tensor with type=DT_VARIANT containing the deduplication data. The tensor is an XLA nested tuple containing N elements (where N is the ratio of the number of embedding to tensor cores per TPU chip). Each element of the nested tuple is a tuple of rank 1 tensors. Each tensor either contains indices (DT_UINT32) for embedding lookup on the TensorCore or weights (DT_FLOAT) to apply to the output of the embedding lookup operation. |
numTables | The number of output activation tensors. If feature descriptor is present in the tpu embedding config, it is equal to the number of features otherwise equal to number of embedding tables in the model. |
config | Serialized TPUEmbeddingConfiguration proto. |
Returns
- a new instance of XlaRecvTPUEmbeddingActivations
public List<Output<Float>> outputs ()
A TensorList of embedding activations containing one Tensor per embedding table in the model.