CbhgLayer<T>
Layers · AiDotNet.NeuralNetworks.Layers
Tacotron’s CBHG module: a 1-D Convolution Bank, Highway network and bidirectional GRU.
Inheritance: LayerBase<T> → CbhgLayer<T>
Implements: IShapeContract
For Beginners
The convolution bank looks at the sequence through windows of many sizes at once, like reading a sentence by its letters, syllables and words together; the GRUs then read the result in both directions.
How It Works
Tacotron (Wang et al. 2017, §3.1, Fig. 2, Table 1): the input is convolved with K sets of 1-D filters of widths
1..K, the outputs are stacked and max-pooled along time (stride 1, width 2), passed through two fixed-width 1-D
projection convolutions, added to the input (residual), fed to a 4-layer highway network, and read by a
bidirectional GRU whose forward and backward outputs are concatenated. Batch normalization is used for every
convolution.
What the paper leaves unstated follows the reference implementation (keithito/tacotron, models/modules.py):
convolutions use “same” padding (for an even width, ⌊(k−1)/2⌋ frames in front and the rest behind), apply their
activation and then batch normalization; the max pooling pads at the end; when the residual sum’s width differs
from the highway width (the post-processing CBHG reads 80 mel bins into 128-wide highways) a linear layer maps it
first; highway gates start biased to −1.
Input [time, channels] or [batch, time, channels]; output [.., time, 2 · gruUnits].
Constructors
| Constructor | Summary |
|---|---|
CbhgLayer(int inputChannels, int bankSize, int bankChannels, int[] projections, int highwayWidth, int gruUnits, CbhgVariant variant, double dropoutRate) |
Creates a CBHG module. |
Properties
| Property | Summary |
|---|---|
bool SupportsTraining { get; } |
|
int OutputChannels { get; } |
Width of the output (both GRU directions). |
Methods
| Method | Summary |
|---|---|
IReadOnlyList<ILayer<T>> GetSubLayers() |
|
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank) |
|
void ResetState() |