tf.tpu.experimental.embedding.serving_embedding_lookup

Apply standard lookup ops with tf.tpu.experimental.embedding configs.

This function is a utility which allows using the tf.tpu.experimental.embedding config objects with standard lookup functions. This can be used when exporting a model which uses tf.tpu.experimental.embedding.TPUEmbedding for serving on CPU. In particular tf.tpu.experimental.embedding.TPUEmbedding only supports lookups on TPUs and should not be part of your serving graph.

Note that TPU specific options (such as max_sequence_length) in the configuration objects will be ignored.

In the following example we take take a trained model (see the documentation for tf.tpu.experimental.embedding.TPUEmbedding for the context) and create a saved model with a serving function that will perform the embedding lookup and pass the results to your model:

model = model_fn(...)
embedding = tf.tpu.experimental.embedding.TPUEmbedding(
    feature_config=feature_config,
    batch_size=1024,
    optimizer=tf.tpu.experimental.embedding.SGD(0.1))
checkpoint = tf.train.Checkpoint(model=model, embedding=embedding)
checkpoint.restore(...)

@tf.function(input_signature=[{'feature_one': tf.TensorSpec(...),
                               'feature_two': tf.TensorSpec(...),
                               'feature_three': tf.TensorSpec(...)}])
def serve_tensors(embedding_featurese):
  embedded_features = tf.tpu.experimental.embedding.serving_embedding_lookup(
      embedding_features, None, embedding.embedding_tables,
      feature_config)
  return model(embedded_features)

model.embedding_api = embedding
tf.saved_model.save(model,
                    export_dir=...,
                    signatures={'serving_default': serve_tensors})

inputs a nested structure of Tensors, SparseTensors or RaggedTensors.