oa::FnMatrix::rnnScan

rnnScan: whole-sequence vanilla-RNN recurrent scan in ONE dispatch. Mathematically identical to running rnnCellLinear for each timestep, but collapses S dispatches into 1. The recurrent weight/bias gradient is computed separately via linearWeightBiasBwd on the saved hPrev.

FnC++ public headers

Function Documentation

RnnScanResult oa::FnMatrix::rnnScan(
	const Matrix & inGatesI,
	const Matrix & inWeightHh,
	oa::I32 inHiddenSize,
	oa::I32 inSeqLen,
	oa::I32 inBatch,
	const Matrix & inBiasHh = Matrix{}
)

Parameters

inGatesI

const Matrix &

[B*S, H] precomputed input projection (row b*S+t = timestep t)

inWeightHh

const Matrix &

—

inHiddenSize

oa::I32

—

inSeqLen

oa::I32

—

inBatch

oa::I32

—

inBiasHh

const Matrix &

Default: Matrix{}

Returns

RnnScanResult

The declared return value.