klyn.math.nn.attention.GatedDeltaAttention.GatedDeltaAttention
constructor
public GatedDeltaAttention(keyHeads as Long, valueHeads as Long, keyDim as Long, valueDim as Long, kernelWidth as Long, epsilon as Double = 1e-6)
Description

Creates an initially zero recurrent state.

Parameters
ParameterDescription
keyHeadsKey head count.
valueHeadsValue count, a multiple of keyHeads.
keyDimKey dimension.
valueDimValue dimension.
kernelWidthCausal convolution width.
epsilonPositive output RMS stabilizer.
Throws
  • ValueException for invalid geometry.
Example

state = GatedDeltaAttention(16l, 32l, 128l, 128l, 4l)