WaveNetResidualBlockLayer<T>
Layers · AiDotNet.NeuralNetworks.Layers
A single WaveNet / Parallel WaveGAN residual block (van den Oord et al. 2016;
Yamamoto et al. 2020) for 1-D waveform/feature data [B, C, T].
Implements the paper’s gated-activation residual unit:
Inheritance: LayerBase<T> → WaveNetResidualBlockLayer<T>
Implements: IShapeContract
How It Works
The dual filter/gate dilated convolutions and the tanh·sigmoid product are
the defining WaveNet gated activation — a plain dilated-conv stack (no gating, no
residual) is NOT WaveNet. The residual connection carries the signal forward
through the deep dilation stack exactly as in the paper’s residual path.
Built from three inner Conv1DLayer instances (filter, gate, 1×1
projection), so the gradient tape and the fused conv kernels are reused — no
hand-written backward. Channel width is constant across the block (the residual
add requires C_out == C_in); the block is reconstructable purely from
(channels, kernelSize, dilation) for Clone/Deserialize.
Constructors
| Constructor | Summary |
|---|---|
WaveNetResidualBlockLayer(int channels, int kernelSize, int dilation) |
Constructs a WaveNet gated-residual block of constant channel width. |
Properties
| Property | Summary |
|---|---|
bool SupportsTraining { get; } |
Methods
| Method | Summary |
|---|---|
IReadOnlyList<ILayer<T>> GetSubLayers() |
|
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank) |
Derived from this type’s [TensorLayout] declarations. |
void ResetState() |
|
void SetTrainingMode(bool isTraining) |
|
void UpdateParameters(T learningRate) |