espnet2.legacy.nets.pytorch_backend.transformer.encoder.Encoder

source

class espnet2.legacy.nets.pytorch_backend.transformer.encoder.Encoder(idim, 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, attention_dropout_rate=0.0, input_layer='conv2d', pos_enc_class=<class 'espnet2.legacy.nets.pytorch_backend.transformer.embedding.PositionalEncoding'>, normalize_before=True, concat_after=False, positionwise_layer_type='linear', positionwise_conv_kernel_size=1, selfattention_layer_type='selfattn', padding_idx=-1, stochastic_depth_rate=0.0, intermediate_layers=None, ctc_softmax=None, conditioning_layer_dim=None)

Bases: Module

Transformer encoder module.

• Parameters:
  • idim (int) – Input dimension.
  • 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 selfattention_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 selfattention_layer_type == “lightconv*” or “dynamiconv*”.
  • conv_usebias (bool) – Whether to use bias in convolution. Only used in selfattention_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.
  • attention_dropout_rate (float) – Dropout rate in attention.
  • input_layer (Union *[*str , torch.nn.Module ]) – Input layer type.
  • 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)
  • positionwise_layer_type (str) – “linear”, “conv1d”, or “conv1d-linear”.
  • positionwise_conv_kernel_size (int) – Kernel size of positionwise conv1d layer.
  • selfattention_layer_type (str) – Encoder attention layer type.
  • padding_idx (int) – Padding idx for input_layer=embed.
  • stochastic_depth_rate (float) – Maximum probability to skip the encoder layer.
  • intermediate_layers (Union *[*List *[*int ] , None ]) – indices of intermediate CTC layer. indices start from 1. if not None, intermediate outputs are returned (which changes return type signature.)

Construct an Encoder object.

forward(xs, masks)

Encode input sequence.

• Parameters:
  • xs (torch.Tensor) – Input tensor (#batch, time, idim).
  • masks (torch.Tensor) – Mask tensor (#batch, 1, time).
• Returns: Output tensor (#batch, time, attention_dim). torch.Tensor: Mask tensor (#batch, 1, time).
• Return type: torch.Tensor

forward_one_step(xs, masks, , cache=None)

Encode input frame.

• Parameters:
  • xs (torch.Tensor) – Input tensor.
  • masks (torch.Tensor) – Mask tensor.
  • cache (List *[*torch.Tensor ]) – List of cache tensors.
• Returns: Output tensor. torch.Tensor: Mask tensor. List[torch.Tensor]: List of new cache tensors.
• Return type: torch.Tensor

get_positionwise_layer(positionwise_layer_type='linear', attention_dim=256, linear_units=2048, dropout_rate=0.1, positionwise_conv_kernel_size=1)

Define positionwise layer.