espnet2.legacy.nets.pytorch_backend.rnn.attentions.AttForward

source

class espnet2.legacy.nets.pytorch_backend.rnn.attentions.AttForward(eprojs, dunits, att_dim, aconv_chans, aconv_filts)

Bases: Module

Forward attention module.

Reference: Forward attention in sequence-to-sequence acoustic modeling for speech synthesis

(https://arxiv.org/pdf/1807.06736.pdf)

• Parameters:
  • eprojs (int) – # projection-units of encoder
  • dunits (int) – # units of decoder
  • att_dim (int) – attention dimension
  • aconv_chans (int) – # channels of attention convolution
  • aconv_filts (int) – filter size of attention convolution

Initialize AttForward.

forward(enc_hs_pad, enc_hs_len, dec_z, att_prev, scaling=1.0, last_attended_idx=None, backward_window=1, forward_window=3)

Calculate AttForward forward propagation.

• Parameters:
  • enc_hs_pad (torch.Tensor) – padded encoder hidden state (B x T_max x D_enc)
  • enc_hs_len (list) – padded encoder hidden state length (B)
  • dec_z (torch.Tensor) – decoder hidden state (B x D_dec)
  • att_prev (torch.Tensor) – attention weights of previous step
  • scaling (float) – scaling parameter before applying softmax
  • last_attended_idx (int) – index of the inputs of the last attended
  • backward_window (int) – backward window size in attention constraint
  • forward_window (int) – forward window size in attetion constraint
• Returns: attention weighted encoder state (B, D_enc)
• Return type: torch.Tensor
• Returns: previous attention weights (B x T_max)
• Return type: torch.Tensor

reset()

Reset states.