espnet.nets.pytorch_backend.transducer.blocks.build_blocks

source

espnet.nets.pytorch_backend.transducer.blocks.build_blocks(net_part: str, idim: int, input_layer_type: str, blocks: List[Dict[str, Any]], repeat_block: int = 0, self_attn_type: str = 'self_attn', positional_encoding_type: str = 'abs_pos', positionwise_layer_type: str = 'linear', positionwise_activation_type: str = 'relu', conv_mod_activation_type: str = 'relu', input_layer_dropout_rate: float = 0.0, input_layer_pos_enc_dropout_rate: float = 0.0, padding_idx: int = -1) → Tuple[Conv2dSubsampling | VGG2L | Sequential, MultiSequential, int, int]

Build custom model blocks.

• Parameters:
  • net_part – Network part, either ‘encoder’ or ‘decoder’.
  • idim – Input dimension.
  • input_layer – Input layer type.
  • blocks – Blocks parameters for network part.
  • repeat_block – Number of times provided blocks are repeated.
  • positional_encoding_type – Positional encoding layer type.
  • positionwise_layer_type – Positionwise layer type.
  • positionwise_activation_type – Positionwise activation type.
  • conv_mod_activation_type – Convolutional module activation type.
  • input_layer_dropout_rate – Dropout rate for input layer.
  • input_layer_pos_enc_dropout_rate – Dropout rate for input layer pos. enc.
  • padding_idx – Padding symbol ID for embedding layer.
• Returns: Input layer all_blocks: Encoder/Decoder network. out_dim: Network output dimension. conv_subsampling_factor: Subsampling factor in frontend CNN.
• Return type: in_layer