Outputs of a tfa.seq2seq.BeamSearchDecoder step.

scores The scores this step, which are the log probabilities over the output vocabulary, possibly penalized by length and attention coverage. When tfa.seq2seq.BeamSearchDecoder is created with output_all_scores=False (default), this will be a float32 Tensor of shape [batch_size, beam_width] containing the top scores corresponding to the predicted IDs. When output_all_scores=True, this contains the scores for all token IDs and has shape [batch_size, beam_width, vocab_size].
predicted_ids The token IDs predicted for this step. A int32 Tensor of shape [batch_size, beam_width].
parent_ids The indices of the parent beam of each beam. A int32 Tensor of shape [batch_size, beam_width].