An op enabling differentiation of TPU Embeddings.

    embedding_variable, sliced_activations, table_id, lookup_id, name=None

This op simply returns its first input, which is assumed to have been sliced from the Tensors returned by TPUEmbeddingDequeueActivations. The presence of this op, and its first argument being a trainable Variable, enables automatic differentiation of graphs containing embeddings via the TPU Embedding Python libraries.


  • embedding_variable: A Tensor of type float32. A trainable variable, enabling optimizers to find this op.
  • sliced_activations: A Tensor of type float32. The embedding activations Tensor to return.
  • table_id: An int that is >= 0. The id of the table in the embedding layer configuration from which these activations were computed.
  • lookup_id: An int that is >= 0. Identifier of the set of embedding indices which produced these activations.
  • name: A name for the operation (optional).


A Tensor of type float32.