fastenhancer_params

Validated FastEnhancer architecture config, official presets and weight mappings; no backend imports.

FASTENHANCER_PRESET_SOURCEconstant
FASTENHANCER_PRESETSconstant
FASTENHANCER_T_ONNXconstantMapping of every weight of build(FastEnhancerParams()) from the pinned FastEnhancer-T ONNX release; every other initializer of the file is a graph constant, li…
MAPPINGSconstantWeight mappings for this family, by name.
FastEnhancerRNNFormerParamsclassRNNFormer stage: per-band GRU over time, then attention across bands.
FastEnhancerParamsclassFolded-inference FastEnhancer config for one spectral frame per call.
FastEnhancerResolvedConfigclassSerializable record of a preset, its overrides and the concrete config.
resolve_fastenhancerfunctionResolve an official preset with validated overrides into a full record.
fastenhancer_weight_shapesfunctionFolded tensors in ONNX/PyTorch layout, keyed by module path.
fastenhancer_mappingfunctionMapping of every weight of build(params) from folded tensors in ONNX export layout.

FASTENHANCER_PRESET_SOURCE = 'aask1357/fastenhancer@e74cab157662dd0d28f381959f9b4870654d5d1f'

FASTENHANCER_PRESETS: Mapping[str, FastEnhancerParams] = {'fastenhancer_t': FastEnhancerParams()}

FASTENHANCER_T_ONNX = fastenhancer_mapping(FASTENHANCER_PRESETS['fastenhancer_t'], 'fastenhancer_t_onnx', SourcePin(uri='https://github.com/aask1357/fastenhancer/releases/download/onnx-vd-v1.0.0/fastenhancer_t.spec.onnx', sha256='915a451f3b1ea8e98c20517c63b50943aa1d540624189f3106cc2e03d09634eb', format='onnx', note='onnx-vd-v1.0.0 FP32 spectral graph (VoiceBank-DEMAND, 16 kHz); code MIT.'), tensor_names={'rf_pre.0.kernel': 'onnx::MatMul_638', 'rf_block.0.rnn.W': 'onnx::GRU_662', 'rf_block.0.rnn.R': 'onnx::GRU_663', 'rf_block.0.rnn.B': 'onnx::GRU_664', 'rf_block.0.rnn_fc.kernel': 'onnx::MatMul_675', 'rf_block.0.attn.qkv.kernel': 'onnx::MatMul_680', 'rf_block.0.attn_fc.kernel': 'onnx::MatMul_702', 'rf_block.1.rnn.W': 'onnx::GRU_724', 'rf_block.1.rnn.R': 'onnx::GRU_725', 'rf_block.1.rnn.B': 'onnx::GRU_726', 'rf_block.1.rnn_fc.kernel': 'onnx::MatMul_737', 'rf_block.1.attn.qkv.kernel': 'onnx::MatMul_742', 'rf_block.1.attn_fc.kernel': 'onnx::MatMul_764', 'rf_post.0.kernel': 'onnx::MatMul_766'}, unused=('/Constant_output_0', '/Constant_1_output_0', '/Constant_2_output_0', '/Constant_3_output_0', '/Constant_4_output_0', '/Constant_6_output_0', '/Constant_7_output_0', '/Constant_8_output_0', '/rf_block.0/Constant_output_0', '/rf_block.0/Constant_1_output_0', '/rf_block.0/attn/Constant_output_0', '/rf_block.0/attn/Constant_1_output_0', '/rf_block.0/attn/Constant_3_output_0', '/rf_block.0/attn/Constant_7_output_0', '/rf_block.0/attn/Constant_11_output_0', '/Constant_14_output_0', '/Constant_17_output_0', '/enc_pre/enc_pre.0/Reshape_1_output_0', '/Concat_output_0', '/rf_block.0/attn/Sqrt_1_output_0', '/Concat_1_output_0', '/Reshape_5_output_0', '/enc_pre/enc_pre.0/Concat_1_output_0', '/enc_pre/enc_pre.0/Concat_2_output_0'))

Mapping of every weight of build(FastEnhancerParams()) from the pinned FastEnhancer-T ONNX release; every other initializer of the file is a graph constant, listed as unused.

MAPPINGS: dict[str, WeightMapping] = {FASTENHANCER_T_ONNX.name: FASTENHANCER_T_ONNX}

Weight mappings for this family, by name.

FastEnhancerRNNFormerParams()

RNNFormer stage: per-band GRU over time, then attention across bands.

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

num_blocks: int = Field(default=2, ge=1)

channels: int = Field(default=20, gt=0)

freq: int = Field(default=16, ge=2)

num_heads: int = Field(default=4, ge=1)

positional_embedding: bool = True

attn_bias: bool = False

post_act: bool = False

FastEnhancerParams()

Folded-inference FastEnhancer config for one spectral frame per call.

BatchNorm and weight normalization are folded into biased layers, matching the released inference graphs; this form is not the trainable architecture. STFT/iSTFT framing belongs to the caller. Changing any field defines a new, untrained architecture unless matching weights exist.

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

family: Literal['fastenhancer'] = 'fastenhancer'

form: Literal['folded_inference'] = 'folded_inference'

n_fft: int = Field(default=512, ge=4)

channels: int = Field(default=24, gt=0)

kernel_size: tuple[int, ...] = (8, 3, 3)

stride: int = Field(default=4, ge=1)

rnnformer: FastEnhancerRNNFormerParams = FastEnhancerRNNFormerParams()

activation: Literal['silu', 'relu'] = 'silu'

mask: Literal['none', 'sigmoid', 'tanh'] = 'none'

input_compression: float = Field(default=0.3, gt=0, le=1, allow_inf_nan=False)

resnet: bool = False

spectral_bins: int

Spectral input/output bins, including the zero-filled Nyquist bin.

input_shape: tuple[int, ...]

The fixed spec_in shape (spectral_bins, 1, 2), without the batch axis.

encoder_bins: int

Frequency positions after the strided input convolution.

FastEnhancerResolvedConfig()

Serializable record of a preset, its overrides and the concrete config.

Reload from params; re-resolving a preset later may differ if the preset table changes. official is true only when the resolved config equals the preset exactly.

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

schema_version: Literal[1] = 1

preset: str

source: str

overrides: dict[str, Any]

official: bool

params: FastEnhancerParams

resolve_fastenhancer(preset: str, overrides: Mapping[str, Any] | None = None) -> FastEnhancerResolvedConfig

Resolve an official preset with validated overrides into a full record.

fastenhancer_weight_shapes(params: FastEnhancerParams) -> dict[str, tuple[int, ...]]

Folded tensors in ONNX/PyTorch layout, keyed by module path.

Conv weights are [out, in, kernel]; *.kernel MatMul weights are [in, out]; rnn.W/rnn.R are [1, 3H, in] and rnn.B is [1, 6H] with ONNX gate order z, r, h; the upsampler is [in, out, kernel].

fastenhancer_mapping(
    params: FastEnhancerParams,
    name: str,
    source: SourcePin,
    *,
    tensor_names: Mapping[str, str] | None = None,
    unused: tuple[str, ...] = (),
) -> WeightMapping

Mapping of every weight of build(params) from folded tensors in ONNX export layout.

Source tensors are keyed by module path, with the shapes fastenhancer_weight_shapes(params) lists; each row checks its source’s shape. GRU tensors must use ONNX gate order z, r, h; PyTorch nn.GRU stores r, z, n, which shapes alone cannot detect.

Parameters

Parameters of fastenhancer_mapping
Name	Type	Default	Description
`params`	`FastEnhancerParams`	Required	The params the tensors were trained with.
`name`	`str`	Required	Mapping name.
`source`	`SourcePin`	Required	The pinned source file.
`tensor_names`	`Mapping[str, str] \| None`	None	Source tensor names for module paths stored under another name.
`unused`	`tuple[str, ...]`	()	Source tensors deliberately not imported.

Returns

Returns of fastenhancer_mapping
Value	Type	Description
`WeightMapping`	`WeightMapping`	The mapping, for ``helia_edge.importers.import_weights``.

Raises

Errors raised by fastenhancer_mapping
Type	Description
`ValueError`	If ``tensor_names`` renames a tensor these params do not have.