espnet.nets.pytorch_backend.transducer.rnn_encoder.RNNP

source

class espnet.nets.pytorch_backend.transducer.rnn_encoder.RNNP(idim: int, rnn_type: str, elayers: int, eunits: int, eprojs: int, subsample: ndarray, dropout_rate: float, aux_output_layers: List = [])

Bases: Module

RNN with projection layer module.

• Parameters:
  • idim – Input dimension.
  • rnn_type – RNNP units type.
  • elayers – Number of RNNP layers.
  • eunits – Number of units ((2 * eunits) if bidirectional).
  • eprojs – Number of projection units.
  • subsample – Subsampling rate per layer.
  • dropout_rate – Dropout rate for RNNP layers.
  • aux_output_layers – Layer IDs for auxiliary RNNP output sequences.

Initialize RNNP module.

forward(rnn_input: Tensor, rnn_len: Tensor, prev_states: List[Tensor] | None = None) → Tuple[Tensor, List[Tensor], Tensor]

RNNP forward.

• Parameters:

  • rnn_input – RNN input sequences. (B, T, D_in)
  • rnn_len – RNN input sequences lengths. (B,)
  • prev_states – RNN hidden states. [N x (B, T, D_proj)]

• Returns: RNN output sequences. (B, T, D_proj) : with or without intermediate RNN output sequences. ((B, T, D_proj), [N x (B, T, D_proj)])

  rnn_len: RNN output sequences lengths. (B,) current_states: RNN hidden states. [N x (B, T, D_proj)]

• Return type: rnn_output