espnet2.asr_transducer.encoder.building.build_conformer_block

source

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

Build Conformer block.

This function constructs a Conformer block based on the provided configuration and main parameters. The Conformer architecture integrates convolutional layers and attention mechanisms to capture both local and global dependencies in the input data.

• Parameters:
  • configuration – A list of dictionaries containing Conformer block configuration settings. Each dictionary should include keys such as “hidden_size”, “linear_size”, “pos_wise_dropout_rate”, and “conv_mod_kernel_size”.
  • main_params – A dictionary containing encoder main parameters, which are used to configure the various components of the Conformer block, including activation functions and normalization settings.
• Returns: A callable that returns a Conformer block instance when invoked.

Examples

>>> config = [{
...     "hidden_size": 256,
...     "linear_size": 512,
...     "pos_wise_dropout_rate": 0.1,
...     "conv_mod_kernel_size": 31,
...     "heads": 4,
...     "att_dropout_rate": 0.1,
...     "dropout_rate": 0.1,
...     "norm_eps": 1e-5
... }]
>>> main_params = {
...     "pos_wise_act": "relu",
...     "conv_mod_act": "relu",
...     "norm_type": "layer_norm",
...     "conv_mod_norm_type": "layer_norm",
...     "dynamic_chunk_training": False
... }
>>> conformer_block = build_conformer_block(config, main_params)()