CifAlignmentLayer<T>

Layers · AiDotNet.NeuralNetworks.Layers

Continuous Integrate-and-Fire (CIF) alignment layer per Gao et al. 2022 “Paraformer” §3.2 / Algorithm 1. Converts a variable-length encoder hidden-state sequence [B, S, D] into a token-aligned acoustic embedding sequence [B, S, D] by predicting per- timestep fire weights and integrating the hidden states until the cumulative weight crosses a unit-mass threshold.

Inheritance: LayerBase<T> → CifAlignmentLayer<T>

Implements: IShapeContract

How It Works

Algorithm (Gao 2022 Algorithm 1):

Output shape — fixed [B, S, D]: the CIF paper’s output length N is data-dependent (depends on round(Σₜ α_t)), which doesn’t fit a static ILayer shape contract. We follow the FunASR runtime convention: declare the output as the same length as the input (a safe upper bound because each α_t ∈ [0, 1] gives at most one fire per step), populate the first predicted_N slots with the CIF tokens, and zero-pad the remainder. Downstream attention layers ignore the padded slots through standard padding-mask handling.

Trainable parameters: only the alpha-predictor’s Dense weights. The integrate-and-fire arithmetic itself is parameter-free and the threshold-crossing is non-differentiable — gradients flow through the alpha predictor only via the upstream loss applied to non-firing accumulation paths. Paraformer’s “alpha scaling” training trick (scaling all α_t so their sum matches the target token count) is the standard way to make the predictor learn alignment in spite of this; consumers that need full alignment supervision should apply that scaling at training time.

Constructors

Constructor Summary
CifAlignmentLayer(int encoderDim, double threshold, double tailThreshold, bool alphaScalingEnabled, double quantityLossWeight) Initializes a new CIF alignment layer.

Properties

Property Summary
T LastQuantityLoss { get; set; } The most recent |Σα − S~|, averaged over the batch, or zero when TargetTokenCount is unset.
Tensor<T> LastPredictedTokenCount { get; set; } The predicted token count from the most recent forward: sum_t alpha_t per batch item.
bool AlphaScalingEnabled { get; set; } Whether the paper’s training-time alpha scaling is applied.
bool SupportsTraining { get; } true: the alpha predictor is genuinely trained.
double QuantityLossWeight { get; set; } Weight λ₂ on the quantity loss.
int? TargetTokenCount { get; set; } Target token count S~ for the current batch, set by the consumer before a training forward pass.

Methods

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