def build_conformer_block(
block: Dict[str, Any],
self_attn_class: str,
pw_layer_type: str,
pw_activation_type: str,
conv_mod_activation_type: str,
) -> ConformerEncoderLayer:
"""Build function for conformer block.
Args:
block: Conformer block parameters.
self_attn_type: Self-attention module type.
pw_layer_type: Positionwise layer type.
pw_activation_type: Positionwise activation type.
conv_mod_activation_type: Convolutional module activation type.
Returns:
: Function to create conformer (encoder) block.
"""
d_hidden = block["d_hidden"]
d_ff = block["d_ff"]
dropout_rate = block.get("dropout-rate", 0.0)
pos_dropout_rate = block.get("pos-dropout-rate", 0.0)
att_dropout_rate = block.get("att-dropout-rate", 0.0)
if pw_layer_type == "linear":
pw_layer = PositionwiseFeedForward(
d_hidden,
d_ff,
pos_dropout_rate,
get_activation(pw_activation_type),
)
else:
raise NotImplementedError("Conformer block only supports linear yet.")
macaron_net = (PositionwiseFeedForward(
d_hidden,
d_ff,
pos_dropout_rate,
get_activation(pw_activation_type),
) if block["macaron_style"] else None)
conv_mod = (ConvolutionModule(
d_hidden,
block["conv_mod_kernel"],
get_activation(conv_mod_activation_type),
) if block["use_conv_mod"] else None)
return lambda: ConformerEncoderLayer(
d_hidden,
self_attn_class(block["heads"], d_hidden, att_dropout_rate),
pw_layer,
macaron_net,
conv_mod,
dropout_rate,
)
def build_conformer_block(
block_arch,
self_attn_class,
pos_enc_class,
pw_layer_type,
pw_activation_type,
conv_mod_activation_type,
):
"""Build function for conformer block.
Args:
block_arch (dict): conformer block parameters
self_attn_type (str): self-attention module type
pos_enc_class (str): positional encoding class
pw_layer_type (str): positionwise layer type
pw_activation_type (str): positionwise activation type
conv_mod_activation_type (str): convolutional module activation type
Returns:
(function): function to create conformer block
"""
d_hidden = block_arch["d_hidden"]
d_ff = block_arch["d_ff"]
heads = block_arch["heads"]
macaron_style = block_arch["macaron_style"]
use_conv_mod = block_arch["use_conv_mod"]
dropout_rate = block_arch[
"dropout-rate"] if "dropout-rate" in block_arch else 0.0
pos_dropout_rate = (block_arch["pos-dropout-rate"]
if "pos-dropout-rate" in block_arch else 0.0)
att_dropout_rate = (block_arch["att-dropout-rate"]
if "att-dropout-rate" in block_arch else 0.0)
if pw_layer_type == "linear":
pw_layer = PositionwiseFeedForward
pw_activation = get_activation(pw_activation_type)
pw_layer_args = (d_hidden, d_ff, pos_dropout_rate, pw_activation)
else:
raise NotImplementedError("Conformer block only supports linear yet.")
if use_conv_mod:
conv_layer = ConvolutionModule
conv_activation = get_activation(conv_mod_activation_type)
conv_layers_args = (d_hidden, block_arch["conv_mod_kernel"],
conv_activation)
return lambda: ConformerEncoderLayer(
d_hidden,
self_attn_class(heads, d_hidden, att_dropout_rate),
pw_layer(*pw_layer_args),
pw_layer(*pw_layer_args) if macaron_style else None,
conv_layer(*conv_layers_args) if use_conv_mod else None,
dropout_rate,
)