ConditionalLayerNormalizationLayer<T>

Layers · AiDotNet.NeuralNetworks.Layers

Layer normalization whose scale and bias are computed from a conditioning vector instead of being free parameters: CLN(x, e) = (e W_γ) ⊙ (x − μ) / σ + e W_β.

Inheritance: LayerBase<T> → ConditionalLayerNormalizationLayer<T>

Implements: IShapeContract

For Beginners

Ordinary layer normalization rescales each feature vector with a learned scale and shift. Here the scale and shift depend on who is speaking, so changing the speaker re-tunes every normalization in the decoder without touching any other weight.

How It Works

AdaSpeech (Chen et al. 2021, §2.2, Eq. 1) replaces every layer normalization in the mel decoder with this layer, conditioned on the speaker embedding E_s: “the conditional network consists of two simple linear layers W_γ and W_β that take speaker embedding E_s as input and output the scale and bias vector respectively”, γ = E_s W_γ, β = E_s W_β. Adapting a voice then fine-tunes only these two matrices (and the speaker embedding), and a deployed voice needs only the resulting γ and β.

Inputs: x is [time, hidden] with a condition [condition] (or [1, condition]), or [batch, time, hidden] with a condition [batch, condition]. Normalization is over the last axis.

Constructors

Constructor Summary
ConditionalLayerNormalizationLayer(int hiddenSize, int conditionSize, double epsilon) Creates a conditional layer normalization.

Properties

Property Summary
BiasFreeLinearLayer<T> BiasProjection { get; } The bias projection W_β.
BiasFreeLinearLayer<T> ScaleProjection { get; } The scale projection W_γ.
bool RequiresMultipleInputs { get; }
bool SupportsTraining { get; }
int ConditionSize { get; } Width of the conditioning vector.
int HiddenSize { get; } Width of the normalized features.

Methods

Method Summary
IReadOnlyList<ILayer<T>> GetSubLayers()
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank) Derived from this type’s [TensorLayout] declarations.
Tensor<T> Forward(Tensor<T> input, Tensor<T> condition) Normalizes input with the scale and bias computed from condition.
void ResetState()