decoder

Decoder architectures for 1-D signal and 2-D spectrogram compression.

build_decoder_2dfunctionBuild a configurable decoder that mirrors the encoder stages.
build_decoder_2d_ssmfunctionDecoder that mixes time with diagonal SSM blocks at every spatial scale.
build_hierarchical_decoder_2dfunctionBuild a coarse + residual/detail decoder for per-level RVQ tensors.
build_hierarchical_adaptor_decoder_2dfunctionBuild a summed-latent decoder with per-level residual adaptors.
build_decoder_2d_spatialfunctionBuild a 2-D spatial decoder that mirrors the spatial encoder.

build_decoder_2d(
    output_len: int = 2048,
    out_ch: int = 1,
    base: int = 32,
    embedding_dim: int = 16,
    multiplier: float = 1.25,
    num_stages: int = 4,
    decoder_block_norm: str = 'none',
    head_norm: str = 'layer',
    use_residual: bool = False,
    activation: str = 'relu',
) -> keras.Model

Build a configurable decoder that mirrors the encoder stages.

Parameters

Parameters of build_decoder_2d
Name	Type	Default	Description
`output_len`	`int`	2048	Number of output time samples.
`out_ch`	`int`	1	Number of output channels.
`base`	`int`	32	Base filter count (mirroring the encoder).
`embedding_dim`	`int`	16	Latent channel dimension.
`multiplier`	`float`	1.25	Filter count multiplier per stage.
`num_stages`	`int`	4	Number of upsample stages (must match encoder).
`decoder_block_norm`	`str`	'none'	Normalization mode for decoder blocks.
`head_norm`	`str`	'layer'	Normalization mode for the output head.
`use_residual`	`bool`	False	If True, add shortcut connections to each stage.
`activation`	`str`	'relu'	Activation function for decoder blocks.

build_decoder_2d_ssm(
    output_len: int = 2048,
    out_ch: int = 1,
    base: int = 32,
    embedding_dim: int = 16,
    multiplier: float = 1.25,
    num_stages: int = 4,
    head_norm: str = 'layer',
    state_size: int = 32,
    num_ssm_blocks: int = 2,
    max_phase: float = 3.141592653589793,
) -> keras.Model

Decoder that mixes time with diagonal SSM blocks at every spatial scale.

Drop-in shape contract for :func:build_decoder_2d: input (B, 1, output_len // 2**num_stages, embedding_dim) output (B, 1, output_len, out_ch)

build_hierarchical_decoder_2d(
    output_len: int = 2048,
    out_ch: int = 1,
    base: int = 32,
    embedding_dim: int = 16,
    num_levels: int = 2,
    multiplier: float = 1.25,
    num_stages: int = 4,
    decoder_block_norm: str = 'none',
    head_norm: str = 'layer',
    use_residual: bool = False,
    activation: str = 'relu',
    detail_scale: float = 0.25,
    include_sum_input: bool = False,
) -> keras.Model

Build a coarse + residual/detail decoder for per-level RVQ tensors.

build_hierarchical_adaptor_decoder_2d(
    output_len: int = 2048,
    out_ch: int = 1,
    base: int = 32,
    embedding_dim: int = 16,
    num_levels: int = 2,
    multiplier: float = 1.25,
    num_stages: int = 4,
    decoder_block_norm: str = 'none',
    head_norm: str = 'layer',
    use_residual: bool = False,
    activation: str = 'relu',
    detail_scale: float = 0.25,
) -> keras.Model

Build a summed-latent decoder with per-level residual adaptors.

build_decoder_2d_spatial(
    output_height: int,
    output_width: int,
    out_ch: int = 2,
    base: int = 32,
    embedding_dim: int = 16,
    multiplier: float = 1.25,
    num_stages: int = 3,
    decoder_block_norm: str = 'none',
    head_norm: str = 'layer',
) -> keras.Model

Build a 2-D spatial decoder that mirrors the spatial encoder.