espnet.nets.pytorch_backend.transformer.decoder.Decoder

source

class espnet.nets.pytorch_backend.transformer.decoder.Decoder(odim, selfattention_layer_type='selfattn', attention_dim=256, attention_heads=4, conv_wshare=4, conv_kernel_length=11, conv_usebias=False, linear_units=2048, num_blocks=6, dropout_rate=0.1, positional_dropout_rate=0.1, self_attention_dropout_rate=0.0, src_attention_dropout_rate=0.0, input_layer='embed', use_output_layer=True, pos_enc_class=<class 'espnet.nets.pytorch_backend.transformer.embedding.PositionalEncoding'>, normalize_before=True, concat_after=False)

Bases: BatchScorerInterface, Module

Transfomer decoder module.

• Parameters:
  • odim (int) – Output diminsion.
  • self_attention_layer_type (str) – Self-attention layer type.
  • attention_dim (int) – Dimension of attention.
  • attention_heads (int) – The number of heads of multi head attention.
  • conv_wshare (int) – The number of kernel of convolution. Only used in self_attention_layer_type == “lightconv*” or “dynamiconv*”.
  • conv_kernel_length (Union *[*int , str ]) – Kernel size str of convolution (e.g. 71_71_71_71_71_71). Only used in self_attention_layer_type == “lightconv*” or “dynamiconv*”.
  • conv_usebias (bool) – Whether to use bias in convolution. Only used in self_attention_layer_type == “lightconv*” or “dynamiconv*”.
  • linear_units (int) – The number of units of position-wise feed forward.
  • num_blocks (int) – The number of decoder blocks.
  • dropout_rate (float) – Dropout rate.
  • positional_dropout_rate (float) – Dropout rate after adding positional encoding.
  • self_attention_dropout_rate (float) – Dropout rate in self-attention.
  • src_attention_dropout_rate (float) – Dropout rate in source-attention.
  • input_layer (Union *[*str , torch.nn.Module ]) – Input layer type.
  • use_output_layer (bool) – Whether to use output layer.
  • pos_enc_class (torch.nn.Module) – Positional encoding module class. PositionalEncoding `or `ScaledPositionalEncoding
  • normalize_before (bool) – Whether to use layer_norm before the first block.
  • concat_after (bool) – Whether to concat attention layer’s input and output. if True, additional linear will be applied. i.e. x -> x + linear(concat(x, att(x))) if False, no additional linear will be applied. i.e. x -> x + att(x)

Construct an Decoder object.

batch_score(ys: Tensor, states: List[Any], xs: Tensor) → Tuple[Tensor, List[Any]]

Score new token batch (required).

• Parameters:
  • ys (torch.Tensor) – torch.int64 prefix tokens (n_batch, ylen).
  • states (List *[*Any ]) – Scorer states for prefix tokens.
  • xs (torch.Tensor) – The encoder feature that generates ys (n_batch, xlen, n_feat).
• Returns: Tuple of : batchfied scores for next token with shape of (n_batch, n_vocab) and next state list for ys.
• Return type: tuple[torch.Tensor, List[Any]]

forward(tgt, tgt_mask, memory, memory_mask)

Forward decoder.

• Parameters:

  • tgt (torch.Tensor) – Input token ids, int64 (#batch, maxlen_out) if input_layer == “embed”. In the other case, input tensor (#batch, maxlen_out, odim).
  • tgt_mask (torch.Tensor) – Input token mask (#batch, maxlen_out). dtype=torch.uint8 in PyTorch 1.2- and dtype=torch.bool in PyTorch 1.2+ (include 1.2).
  • memory (torch.Tensor) – Encoded memory, float32 (#batch, maxlen_in, feat).
  • memory_mask (torch.Tensor) – Encoded memory mask (#batch, maxlen_in). dtype=torch.uint8 in PyTorch 1.2- and dtype=torch.bool in PyTorch 1.2+ (include 1.2).

• Returns: Decoded token score before softmax (#batch, maxlen_out, odim) : if use_output_layer is True. In the other case,final block outputs (#batch, maxlen_out, attention_dim).

  torch.Tensor: Score mask before softmax (#batch, maxlen_out).

• Return type: torch.Tensor

forward_one_step(tgt, tgt_mask, memory, *, cache=None)

Forward one step.

• Parameters:
  • tgt (torch.Tensor) – Input token ids, int64 (#batch, maxlen_out).
  • tgt_mask (torch.Tensor) – Input token mask (#batch, maxlen_out). dtype=torch.uint8 in PyTorch 1.2- and dtype=torch.bool in PyTorch 1.2+ (include 1.2).
  • memory (torch.Tensor) – Encoded memory, float32 (#batch, maxlen_in, feat).
  • cache (List *[*torch.Tensor ]) – List of cached tensors. Each tensor shape should be (#batch, maxlen_out - 1, size).
• Returns: Output tensor (batch, maxlen_out, odim). List[torch.Tensor]: List of cache tensors of each decoder layer.
• Return type: torch.Tensor

score(ys, state, x)

Score.