tfa.seq2seq.BeamSearchDecoderState

State of a tfa.seq2seq.BeamSearchDecoder.

cell_state The cell state returned at the previous time step.
log_probs The accumulated log probabilities of each beam. A float32 Tensor of shape [batch_size, beam_width].
finished The finished status of each beam. A bool Tensor of shape [batch_size, beam_width].
lengths The accumulated length of each beam. An int64 Tensor of shape [batch_size, beam_width].
accumulated_attention_probs

accumulated_attention_prob Accumulation of the attention probabilities (used to compute the coverage penalty)