conformer_params

Typed parameters of the conformer family; importable without Keras.

SubsampleBlockParams()

Subsample block parameters

model_config = ConfigDict(frozen=True, extra='forbid')

depth: int = 256

Depth

kernel_size: int = 3

Kernel size

strides: int = 2

Stride size

ConformerBlockParams()

Conformer block parameters

model_config = ConfigDict(frozen=True, extra='forbid')

depth: int = 256

Depth

fc_ex_factor: float = 4

FC expansion factor

fc_res_factor: float = 0.5

FC residual factor

embedding: str = 'relative'

Embedding type

num_heads: int = 4

Number of heads

kernel_size: int = 9

Kernel size

dropout: float = 0.1

Dropout rate

use_bias: bool = True

Use bias

ConformerParams()

Conformer parameters

model_config = ConfigDict(frozen=True, extra='forbid')

family: Literal['conformer'] = 'conformer'

Model family

subsamples: list[SubsampleBlockParams] = Field(default_factory=list, description='Subsample blocks')

Subsample blocks

blocks: list[ConformerBlockParams] = Field(default_factory=list, description='Conformer blocks')

Conformer blocks

output_activation: str | None = Field(default=None, description='Output activation')

Output activation

include_top: bool = Field(default=True, description='Include top')

Include top

num_classes: int | None = Field(default=None, gt=0, description='Classes of the output layer')

Classes of the output layer