CitrinetBlockLayer<T>
Layers · AiDotNet.NeuralNetworks.Layers
A single NVIDIA Citrinet residual mega-block over [B, C, T] feature data
(Majumdar et al., 2021, “Citrinet: Closing the Gap between Non-Autoregressive and
Autoregressive End-to-End Models for ASR”, https://arxiv.org/abs/2104.01721).
Inheritance: LayerBase<T> → CitrinetBlockLayer<T>
Implements: IShapeContract
How It Works
Implements the paper’s block exactly:
The 1-D time-channel separable convolution (depthwise-temporal + pointwise-channel), the
squeeze-and-excitation channel attention, and the residual connection are Citrinet’s three
defining features — a plain 1-D conv stack (no separability, no SE, no residual) is NOT
Citrinet. The residual path carries the signal through the deep block stack so gradients flow;
channel width C is constant across the block (the residual add requires C_out == C_in).
Composed from tape-aware inner layers (DepthwiseConv1DLayer,
Conv1DLayer, BatchNormalizationLayer,
SqueezeAndExcitationLayer) and IEngine ops, so the gradient tape
autodiffs the block — no hand-written backward. The block is fully reconstructable from
(channels, kernelSize, numSubBlocks, seReductionRatio, dropoutRate, stride) for
Clone/Deserialize.
Constructors
| Constructor | Summary |
|---|---|
CitrinetBlockLayer(int channels, int kernelSize, int numSubBlocks, int seReductionRatio, double dropoutRate, int stride, int seed) |
Constructs a Citrinet residual mega-block. |
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) |