nexuml_library.scenarios.model.conv_ae

nexuml_library.scenarios.model.conv_ae

Convolutional autoencoder model scenario fragments.

spectrogram_conv_ae

spectrogram_conv_ae(
    input_shape: tuple[int, int, int] = (1, 64, 64),
    latent_dim: int = 64,
    channel_schedule: list[int] | None = None,
    activation: str = "relu",
    score_reduction: str = "mean",
) -> PipelineSpec

2D convolutional autoencoder with reconstruction loss and anomaly score.

Returns:

Name	Type	Description
PipelineSpec	PipelineSpec	Pipeline with encoder, decoder, reconstruction loss and anomaly score layers.

spectrogram_conv_cvae

spectrogram_conv_cvae(
    input_shape: tuple[int, int, int] = (1, 64, 64),
    encoder_dim: int = 128,
    latent_dim: int = 32,
    beta: float = 1.0,
    channel_schedule: list[int] | None = None,
    activation: str = "relu",
    score_reduction: str = "mean",
) -> PipelineSpec

2D convolutional variational autoencoder with anomaly score.

Returns:

Name	Type	Description
PipelineSpec	PipelineSpec	Pipeline with encoder, variational latent, decoder, reconstruction loss and anomaly score layers.

conv_ae_lmbe

conv_ae_lmbe(
    sample_rate: int = 16000,
    n_mels: int = 128,
    n_fft: int = 1024,
    hop_length: int = 512,
    time_frames: int = 128,
    latent_dim: int = 64,
    channel_schedule: list[int] | None = None,
    activation: str = "relu",
) -> PipelineSpec

Waveform -> LMBE -> convolutional AE.

Returns:

Name	Type	Description
PipelineSpec	PipelineSpec	Full waveform-to-reconstruction pipeline with LMBE feature extraction and convolutional autoencoder.

conv_cvae_lmbe

conv_cvae_lmbe(
    sample_rate: int = 16000,
    n_mels: int = 128,
    n_fft: int = 1024,
    hop_length: int = 512,
    time_frames: int = 128,
    encoder_dim: int = 128,
    latent_dim: int = 32,
    beta: float = 1.0,
    channel_schedule: list[int] | None = None,
    activation: str = "relu",
) -> PipelineSpec

Waveform -> LMBE -> convolutional CVAE.

Returns:

Name	Type	Description
PipelineSpec	PipelineSpec	Full waveform-to-reconstruction pipeline with LMBE feature extraction and convolutional variational autoencoder.