nexuml_library.scenarios.asd.audioset_conv_ae_clshead¶
Convolutional autoencoder on AudioSet with a latent classification head.
audioset_conv_ae_clshead(
data_root: str = "audioset_hf/full",
download: bool = False,
sample_rate: int = 16000,
clip_num_samples: int = 160000,
n_mels: int = 128,
hop_length: int = 512,
latent_dim: int = 64,
channel_schedule: list[int] | None = None,
num_classes: int = 527,
lr: float = 0.001,
batch_size: int = 128,
max_epochs: int = 10,
validate_layout: bool = False,
) -> ScenarioSpec
Conv AE trained on AudioSet waveforms with reconstruction + classification.
A LatentClassificationHead is added on top of the encoder latent space,
and both reconstruction and classification losses are optimized jointly.
Returns:
| Name | Type | Description |
|---|---|---|
ScenarioSpec |
ScenarioSpec
|
Assembled scenario with pipeline, training, data and evaluation. |