espnet.nets.pytorch_backend.rnn.attentions.AttMultiHeadAdd

source

class espnet.nets.pytorch_backend.rnn.attentions.AttMultiHeadAdd(eprojs, dunits, aheads, att_dim_k, att_dim_v, han_mode=False)

Bases: Module

Multi head additive attention

Reference: Attention is all you need : (https://arxiv.org/abs/1706.03762)

This attention is multi head attention using additive attention for each head.

• Parameters:
  • eprojs (int) – # projection-units of encoder
  • dunits (int) – # units of decoder
  • aheads (int) – # heads of multi head attention
  • att_dim_k (int) – dimension k in multi head attention
  • att_dim_v (int) – dimension v in multi head attention
  • han_mode (bool) – flag to swith on mode of hierarchical attention and not store pre_compute_k and pre_compute_v

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(enc_hs_pad, enc_hs_len, dec_z, att_prev, **kwargs)

AttMultiHeadAdd forward

• 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) – dummy (does not use)
• Returns: attention weighted encoder state (B, D_enc)
• Return type: torch.Tensor
• Returns: list of previous attention weight (B x T_max) * aheads
• Return type: list

reset()

reset states