oa::FnLoss

FnLoss namespace in the OA Ml public surface.

NsC++ public headers

Public Functions

Matrix oa::FnLoss::bce(const Matrix & inA, const Matrix & inB)

Matrix oa::FnLoss::bceBwd(const Matrix & inA, const Matrix & inB)

oa::Matrix oa::FnLoss::crossEntropy(const oa::Matrix & inLogits, const oa::Matrix & inTargets)

Matrix oa::FnLoss::crossEntropyBwd(const Matrix & inLogits, const Matrix & inTargets)

DqnLossResult oa::FnLoss::dqn(const oa::Matrix & inQ, const oa::Matrix & inAction, const oa::Matrix & inReward, const oa::Matrix & inNextQ, const oa::Matrix & inTerminated, const oa::Matrix & inTruncated, const DqnLossConfig & inConfig = {})

Matrix oa::FnLoss::l1(const Matrix & inA, const Matrix & inB)

Matrix oa::FnLoss::l1Bwd(const Matrix & inA, const Matrix & inB)

const char * oa::FnLoss::lastName()

Matrix oa::FnLoss::maskedCrossEntropy(const Matrix & inLogits, const Matrix & inTargets, const Matrix & inMask, oa::I32 inValidCount)

Matrix oa::FnLoss::maskedCrossEntropyBwd(const Matrix & inLogits, const Matrix & inTargets, const Matrix & inMask, oa::I32 inValidCount)

Matrix oa::FnLoss::mse(const Matrix & inA, const Matrix & inB)

Matrix oa::FnLoss::mseBwd(const Matrix & inA, const Matrix & inB)

PpoLossResult oa::FnLoss::ppo(const oa::Matrix & inNewLogProbability, const oa::Matrix & inOldLogProbability, const oa::Matrix & inAdvantage, const oa::Matrix & inValue, const oa::Matrix & inTargetReturn, const oa::Matrix & inEntropy, const PpoLossConfig & inConfig = {})

oa::Matrix oa::FnLoss::ppoClippedPolicy(const oa::Matrix & inNewLogProbability, const oa::Matrix & inOldLogProbability, const oa::Matrix & inAdvantage, oa::F32 inClipEpsilon = 0.2F)

oa::Matrix oa::FnLoss::ppoClippedPolicyBwd(const oa::Matrix & inNewLogProbability, const oa::Matrix & inOldLogProbability, const oa::Matrix & inAdvantage, oa::F32 inClipEpsilon = 0.2F)

oa::Matrix oa::FnLoss::sacActor(const oa::Matrix & inQ1, const oa::Matrix & inQ2, const oa::Matrix & inLogProbability, oa::F32 inEntropyCoefficient = 0.2F)

SacCriticLossResult oa::FnLoss::sacCritic(const oa::Matrix & inQ1, const oa::Matrix & inQ2, const oa::Matrix & inReward, const oa::Matrix & inNextQ1, const oa::Matrix & inNextQ2, const oa::Matrix & inNextLogProbability, const oa::Matrix & inTerminated, const oa::Matrix & inTruncated, const SacLossConfig & inConfig = {})

void oa::FnLoss::setLastName(const char * inName)

Matrix oa::FnLoss::smoothL1(const Matrix & inA, const Matrix & inB)

Matrix oa::FnLoss::smoothL1Bwd(const Matrix & inA, const Matrix & inB)

Function Documentation

Matrix oa::FnLoss::bce(
	const Matrix & inA,
	const Matrix & inB
)

bce: binary cross-entropy. -(b*log(a) + (1-b)*log(1-a)), clamped for stability.

Parameters

inA

const Matrix &

—

inB

const Matrix &

—

Returns

Matrix

The declared return value.

Matrix oa::FnLoss::bceBwd(
	const Matrix & inA,
	const Matrix & inB
)

bceBwd: gradient w.r.t. inA: (a-b)/(a*(1-a))/N.

Parameters

inA

const Matrix &

—

inB

const Matrix &

—

Returns

Matrix

The declared return value.

oa::Matrix oa::FnLoss::crossEntropy(
	const oa::Matrix & inLogits,
	const oa::Matrix & inTargets
)

crossEntropy: Mean cross-entropy over rank-two logits and UInt8, UInt32, or non-negative Int32 class-index targets; the scalar result is Float32, and an out-of-range target produces NaN without an out-of-bounds read.

Parameters

inLogits

const oa::Matrix &

—

inTargets

const oa::Matrix &

—

Returns

oa::Matrix

The declared return value.

Matrix oa::FnLoss::crossEntropyBwd(
	const Matrix & inLogits,
	const Matrix & inTargets
)

crossEntropyBwd: gradient w.r.t. logits: (softmax(logits) - onehot(targets)) / batch

Parameters

inLogits

const Matrix &

[batch, classes] unnormalized logits

inTargets

const Matrix &

[batch] class indices (UInt8, UInt32, or non-negative Int32)

Returns

Matrix

The declared return value.

DqnLossResult oa::FnLoss::dqn(
	const oa::Matrix & inQ,
	const oa::Matrix & inAction,
	const oa::Matrix & inReward,
	const oa::Matrix & inNextQ,
	const oa::Matrix & inTerminated,
	const oa::Matrix & inTruncated,
	const DqnLossConfig & inConfig = {}
)

No public source comment is attached to this declaration.

Parameters

inQ

const oa::Matrix &

—

inAction

const oa::Matrix &

—

inReward

const oa::Matrix &

—

inNextQ

const oa::Matrix &

—

inTerminated

const oa::Matrix &

—

inTruncated

const oa::Matrix &

—

inConfig

const DqnLossConfig &

Default: {}

Returns

DqnLossResult

The declared return value.

Matrix oa::FnLoss::l1(
	const Matrix & inA,
	const Matrix & inB
)

l1: mean absolute error. mean(|a-b|).

Parameters

inA

const Matrix &

—

inB

const Matrix &

—

Returns

Matrix

The declared return value.

Matrix oa::FnLoss::l1Bwd(
	const Matrix & inA,
	const Matrix & inB
)

l1Bwd: gradient w.r.t. inA: sign(a-b)/N.

Parameters

inA

const Matrix &

—

inB

const Matrix &

—

Returns

Matrix

The declared return value.

const char * oa::FnLoss::lastName()

Returns the name of the most recently called loss function, or nullptr.

Returns

const char *

The declared return value.

Matrix oa::FnLoss::maskedCrossEntropy(
	const Matrix & inLogits,
	const Matrix & inTargets,
	const Matrix & inMask,
	oa::I32 inValidCount
)

maskedCrossEntropy: cross-entropy over only rows where inMask is non-zero.

Parameters

inLogits

const Matrix &

—

inTargets

const Matrix &

—

inMask

const Matrix &

—

inValidCount

oa::I32

—

Returns

Matrix

The declared return value.

Matrix oa::FnLoss::maskedCrossEntropyBwd(
	const Matrix & inLogits,
	const Matrix & inTargets,
	const Matrix & inMask,
	oa::I32 inValidCount
)

maskedCrossEntropyBwd: backward for maskedCrossEntropy.

Parameters

inLogits

const Matrix &

—

inTargets

const Matrix &

—

inMask

const Matrix &

—

inValidCount

oa::I32

—

Returns

Matrix

The declared return value.

Matrix oa::FnLoss::mse(
	const Matrix & inA,
	const Matrix & inB
)

mse: mean squared error. mean((a-b)^2).

Parameters

inA

const Matrix &

—

inB

const Matrix &

—

Returns

Matrix

The declared return value.

Matrix oa::FnLoss::mseBwd(
	const Matrix & inA,
	const Matrix & inB
)

mseBwd: gradient w.r.t. inA: 2*(a-b)/N.

Parameters

inA

const Matrix &

—

inB

const Matrix &

—

Returns

Matrix

The declared return value.

PpoLossResult oa::FnLoss::ppo(
	const oa::Matrix & inNewLogProbability,
	const oa::Matrix & inOldLogProbability,
	const oa::Matrix & inAdvantage,
	const oa::Matrix & inValue,
	const oa::Matrix & inTargetReturn,
	const oa::Matrix & inEntropy,
	const PpoLossConfig & inConfig = {}
)

No public source comment is attached to this declaration.

Parameters

inNewLogProbability

const oa::Matrix &

—

inOldLogProbability

const oa::Matrix &

—

inAdvantage

const oa::Matrix &

—

inValue

const oa::Matrix &

—

inTargetReturn

const oa::Matrix &

—

inEntropy

const oa::Matrix &

—

inConfig

const PpoLossConfig &

Default: {}

Returns

PpoLossResult

The declared return value.

oa::Matrix oa::FnLoss::ppoClippedPolicy(
	const oa::Matrix & inNewLogProbability,
	const oa::Matrix & inOldLogProbability,
	const oa::Matrix & inAdvantage,
	oa::F32 inClipEpsilon = 0.2F
)

No public source comment is attached to this declaration.

Parameters

inNewLogProbability

const oa::Matrix &

—

inOldLogProbability

const oa::Matrix &

—

inAdvantage

const oa::Matrix &

—

inClipEpsilon

oa::F32

Default: 0.2F

Returns

oa::Matrix

The declared return value.

oa::Matrix oa::FnLoss::ppoClippedPolicyBwd(
	const oa::Matrix & inNewLogProbability,
	const oa::Matrix & inOldLogProbability,
	const oa::Matrix & inAdvantage,
	oa::F32 inClipEpsilon = 0.2F
)

No public source comment is attached to this declaration.

Parameters

inNewLogProbability

const oa::Matrix &

—

inOldLogProbability

const oa::Matrix &

—

inAdvantage

const oa::Matrix &

—

inClipEpsilon

oa::F32

Default: 0.2F

Returns

oa::Matrix

The declared return value.

oa::Matrix oa::FnLoss::sacActor(
	const oa::Matrix & inQ1,
	const oa::Matrix & inQ2,
	const oa::Matrix & inLogProbability,
	oa::F32 inEntropyCoefficient = 0.2F
)

No public source comment is attached to this declaration.

Parameters

inQ1

const oa::Matrix &

—

inQ2

const oa::Matrix &

—

inLogProbability

const oa::Matrix &

—

inEntropyCoefficient

oa::F32

Default: 0.2F

Returns

oa::Matrix

The declared return value.

SacCriticLossResult oa::FnLoss::sacCritic(
	const oa::Matrix & inQ1,
	const oa::Matrix & inQ2,
	const oa::Matrix & inReward,
	const oa::Matrix & inNextQ1,
	const oa::Matrix & inNextQ2,
	const oa::Matrix & inNextLogProbability,
	const oa::Matrix & inTerminated,
	const oa::Matrix & inTruncated,
	const SacLossConfig & inConfig = {}
)

No public source comment is attached to this declaration.

Parameters

inQ1

const oa::Matrix &

—

inQ2

const oa::Matrix &

—

inReward

const oa::Matrix &

—

inNextQ1

const oa::Matrix &

—

inNextQ2

const oa::Matrix &

—

inNextLogProbability

const oa::Matrix &

—

inTerminated

const oa::Matrix &

—

inTruncated

const oa::Matrix &

—

inConfig

const SacLossConfig &

Default: {}

Returns

SacCriticLossResult

The declared return value.

void oa::FnLoss::setLastName(
	const char * inName
)

Internal: called by each loss function to record its name.

Parameters

inName

const char *

—

Returns

void

The declared return value.

Matrix oa::FnLoss::smoothL1(
	const Matrix & inA,
	const Matrix & inB
)

smoothL1: smooth L1 / Huber loss (beta=1.0). mean over all elements.

Parameters

inA

const Matrix &

—

inB

const Matrix &

—

Returns

Matrix

The declared return value.

Matrix oa::FnLoss::smoothL1Bwd(
	const Matrix & inA,
	const Matrix & inB
)

smoothL1Bwd: gradient w.r.t. inA.

Parameters

inA

const Matrix &

—

inB

const Matrix &

—

Returns

Matrix

The declared return value.