SpectrogramStripeDropLayer<T>
Layers · AiDotNet.NeuralNetworks.Layers
SpecAugment-style stripe dropping on a spectrogram during training (Park et al. 2019), as PANNs and HTS-AT
apply it inside the network (torchlibrosa SpecAugmentation): zeroes random bands of time frames and then
random bands of frequency bins in every training forward, and passes the input through unchanged otherwise.
Inheritance: LayerBase<T> → SpectrogramStripeDropLayer<T>
Implements: IShapeContract
For Beginners
While training, a few random slices of time and of pitch are blanked out, so the model learns not to depend on any single moment or frequency band.
How It Works
Input: [batch, channels, time, frequency]. For each sample, timeStripes times: a width
w is drawn from [0, timeDropWidth), a start from [0, T − w), and those frames are zeroed in
every channel; then the same along frequency with freqDropWidth and freqStripes. HTS-AT uses
widths 64 (time) and 8 (frequency) with two stripes each. A width is capped below the axis length, so a short
spectrogram is never wiped out entirely.
The drop is a multiplication by a 0/1 mask, so gradients flow through the kept positions.
Constructors
| Constructor | Summary |
|---|---|
SpectrogramStripeDropLayer(int timeDropWidth, int timeStripes, int freqDropWidth, int freqStripes) |
Creates the stripe dropper. |
Properties
| Property | Summary |
|---|---|
bool SupportsTraining { get; } |
Methods
| Method | Summary |
|---|---|
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank) |
Derived from this type’s [TensorLayout] declarations. |
void ResetState() |