DilatedResidualConvBlock<T>

Layers · AiDotNet.NeuralNetworks.Layers

SpeedySpeech’s residual block: n repetitions of a dilated 1D convolution, ReLU and temporal batch normalization, with a residual connection around them: y = x + f(x).

Inheritance: LayerBase<T> → DilatedResidualConvBlock<T>

Implements: IShapeContract

For Beginners

Each convolution looks at a window of neighbouring frames, spaced further apart as the dilation grows, so a deep stack can see a long stretch of the sequence without attention.

How It Works

SpeedySpeech (Vainer & Dušek 2020, §3.2): “progressively dilated residual convolutional blocks, each of which contains a 1D convolution, ReLU activation and temporal batch normalization. A residual connection is applied.” The kernel size and the exact composition are not in the paper; they follow the authors’ implementation (github.com/janvainer/speedyspeech, layers.ResidualBlock): each convolution is unpadded, its output is zero-padded back to the input length (⌊d(k−1)/2⌋ frames in front, the rest behind, so even kernels work), then ReLU, then batch normalization of each channel over the batch and time axes. The paper’s “26 encoder blocks with dilations 1, 1, 2, 2, 4, 4” are 13 of these blocks with two convolutions each.

Input and output are [time, channels] or [batch, time, channels].

Constructors

Constructor Summary
DilatedResidualConvBlock(int channels, int kernelSize, int dilation, int convolutions) Creates a residual block.

Properties

Property Summary
bool SupportsTraining { get; }
int Dilation { get; } Dilation of the convolutions.
int KernelSize { get; } Kernel of the convolutions.

Methods

Method Summary
IReadOnlyList<ILayer<T>> GetSubLayers()
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank) Derived from this type’s [TensorLayout] declarations.
void ResetState()