GatedConvolutionBlockLayer<T>

Layers · AiDotNet.NeuralNetworks.Layers

Deep Voice 3’s convolution block: dropout, a 1-D convolution to twice the channels, a gated linear unit, and an optional residual connection scaled by √0.5.

Inheritance: LayerBase<T> → GatedConvolutionBlockLayer<T>

Implements: IShapeContract

For Beginners

A convolution whose output decides, channel by channel, how much of itself to let through, which keeps deep stacks of convolutions trainable.

How It Works

Ping et al. 2018, §3.3, Fig. 2: “a 1-D convolution filter, a gated-linear unit as a learnable nonlinearity, a residual connection to the input, and a scaling factor of √0.5”; the convolution output of 2c channels splits into the input half and the gate half, a ⊙ σ(b). Causal convolutions pad k − 1 steps on the left, non-causal ones (k − 1)/2 on both sides; dropout is applied to the input. Weights follow the convolutional sequence-to-sequence initialization the paper cites (Gehring et al. 2017; reference implementation r9y9/deepvoice3_pytorch): weight normalization with directions drawn from N(0, std_mul (1 − p) / (k · c_in)).

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

Constructors

Constructor Summary
GatedConvolutionBlockLayer(int inChannels, int outChannels, int kernelSize, int dilation, bool causal, double dropoutRate, bool residual, double stdMultiplier) Creates the block.

Properties

Property Summary
bool SupportsTraining { get; }

Methods

Method Summary
IReadOnlyList<ILayer<T>> GetSubLayers()
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank)
void ResetState()