asr_model

• encoder

  input：
  speech（16，80，183）# 183属于batch中最大元素决定 speech_length
  text （16，6）# 6由batch最大值决定 text_length
  

  • make_pad_mask

    mask ：（16,183）
    

  • subsampling

    input(speech,mask)
    

    • conv(speech)

      torch.nn.Conv2d(1, odim, 3, 2),
      torch.nn.ReLU(),
      torch.nn.Conv2d(odim, odim, 3, 2),
      torch.nn.ReLU()
      # output (16,256,45,19)
      

    • self.out: linear

      torch.nn.Linear(odim * (((idim - 1) // 2 - 1) // 2), odim))
      # output (16,45，256)
      

    • self.pos_enc

      pos_emb (1,45,256) 
      # output  
      # speech = torch.nn.Dropout(speech)  (16,45，256)
      # pos_emb = torch.nn.Dropout(pos_emb) (1,45,256) 
      

  • $subsampling

    #output: speech , pos_emb , mask(16,1,45)  #x_mask[:, :, :-2:2][:, :, :-2:2]
    

  • add_optional_chunk_mask

    add_optional_chunk_mask