espnet2.asr_transducer.encoder.building.build_branchformer_block

source

espnet2.asr_transducer.encoder.building.build_branchformer_block(configuration: List[Dict[str, Any]], main_params: Dict[str, Any]) → Branchformer

Build Branchformer block.

This function constructs a Branchformer block, which is a component of the encoder architecture in the Transducer model. The Branchformer block leverages attention mechanisms and convolutional layers to process input data efficiently.

• Parameters:

  • configuration –

    A list of dictionaries containing the configuration for the Branchformer block. Each dictionary must include the keys:

    • hidden_size: Size of the hidden layer.
    • linear_size: Size of the linear layer.
    • conv_mod_kernel_size: Kernel size for the convolutional

    module.

    • dropout_rate: Dropout rate for the block.
    • heads: Number of attention heads (optional, default is 4).
    • att_dropout_rate: Dropout rate for attention (optional).
    • norm_eps: Epsilon value for normalization (optional).
    • norm_partial: Partial value for normalization (optional).

  • main_params –

    A dictionary containing the main parameters for the encoder, including:

    • conv_mod_norm_type: Type of normalization for the

    convolution module.

    • simplified_att_score: Boolean indicating if simplified attention scoring is used.
• Returns: A callable function that returns a Branchformer block : when invoked.
• Return type:Branchformer

Examples

>>> config = [
...     {
...         "hidden_size": 256,
...         "linear_size": 128,
...         "conv_mod_kernel_size": 3,
...         "dropout_rate": 0.1,
...         "heads": 4
...     }
... ]
>>> main_params = {
...     "conv_mod_norm_type": "layer_norm",
...     "simplified_att_score": False
... }
>>> branchformer_block = build_branchformer_block(config, main_params)
>>> block = branchformer_block()  # Instantiate the block