espnet2.legacy.nets.pytorch_backend.transducer.transformer_decoder_layer.TransformerDecoderLayer

source

class espnet2.legacy.nets.pytorch_backend.transducer.transformer_decoder_layer.TransformerDecoderLayer(hdim: int, self_attention: MultiHeadedAttention, feed_forward: PositionwiseFeedForward, dropout_rate: float)

Bases: Module

Transformer decoder layer module for custom Transducer model.

• Parameters:
  • hdim – Hidden dimension.
  • self_attention – Self-attention module.
  • feed_forward – Feed forward module.
  • dropout_rate – Dropout rate.

Construct an DecoderLayer object.

forward(sequence: Tensor, mask: Tensor, cache: Tensor | None = None)

Compute previous decoder output sequences.

• Parameters:
  • sequence – Transformer input sequences. (B, U, D_dec)
  • mask – Transformer intput mask sequences. (B, U)
  • cache – Cached decoder output sequences. (B, (U - 1), D_dec)
• Returns: Transformer output sequences. (B, U, D_dec) mask: Transformer output mask sequences. (B, U)
• Return type: sequence