ContextNetBlockLayer<T>

Layers ยท AiDotNet.NeuralNetworks.Layers

A paper-faithful ContextNet convolution block over channels-first [B, C, T] features.

Inheritance: LayerBase<T> โ†’ ContextNetBlockLayer<T>

Implements: IShapeContract

How It Works

Han et al. (INTERSPEECH 2020, arXiv:2005.03191, sections 2.2.1-2.2.4) define each convolution as depthwise-separable Conv1D followed by batch normalization and Swish. The stacked convolution path is recalibrated by utterance-wide squeeze-and-excitation, added to a projected residual, and passed through Swish. C0 and C22 retain the main path and SE but omit the residual projection, as specified by Table 1.

Constructors

Constructor Summary
ContextNetBlockLayer(int inputChannels, int outputChannels, int kernelSize, int numConvolutions, int seReductionRatio, double dropoutRate, int stride, bool useResidual, int seed)

Properties

Property Summary
bool SupportsTraining { get; }

Methods

Method Summary
IReadOnlyList<ILayer<T>> GetSubLayers()
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank)
void ResetState()
void SetTrainingMode(bool isTraining)
void UpdateParameters(T learningRate)