tf.compat.v1.nn.embedding_lookup_sparse