Informations de référence sur la fonction complète BrainScript

Article
03/09/2017

Cette section fournit des informations sur les fonctions intégrées BrainScript.

Les déclarations de toutes les fonctions intégrées se trouvent dans la CNTK.core.bs située en regard du binaire CNTK.

Les opérations primitives et les couches sont déclarées dans l’espace de noms global. Des opérations supplémentaires sont déclarées dans les espaces de noms et sont fournies avec le préfixe respectif (par exemple, BS.RNN.LSTMP).

Couches

DenseLayer {outDim, bias= true, activation=Identity, init='uniform', initValueScale=1}
ConvolutionalLayer {numOutputChannels, filterShape, activation = Identity,
init = "uniform", initValueScale = 1,
stride = 1, pad = false, lowerPad = 0, upperPad = 0, bias=true}
MaxPoolingLayer {filterShape, stride = 1, pad = false, lowerPad = 0, upperPad = 0}
AveragePoolingLayer {filterShape, stride = 1, pad = false, lowerPad = 0, upperPad = 0}
EmbeddingLayer {outDim, embeddingPath = '', transpose = false}
RecurrentLSTMLayer {outputDim, cellShape = None, goBackwards = false, enableSelfStabilization = false}
DelayLayer {T=1, defaultHiddenActivation=0}
Dropout
BatchNormalizationLayer {spatialRank = 0, initialScale = 1, normalizationTimeConstant = 0, blendTimeConstant = 0, epsilon = 0.00001, useCntkEngine = true}
LayerNormalizationLayer {initialScale = 1, initialBias = 0}
StabilizerLayer{}
FeatureMVNLayer{}

Génération de couches

Sequential()

Fonctions d’activation

Sigmoid(x)
Tanh(x)
ReLU(x)
Softmax(x, axis=0)
LogSoftmax(x)
Hardmax(x)

Opérations au niveau des éléments, unaire

Abs (x)
Ceil (x)
Cosine (x)
Clip (x, minValue, maxValue)
Exp (x)
Floor (x)
Log (x)
Negate (x)
-x
BS.Boolean.Not (b)
!x
Reciprocal (x)
Round (x)
Sin (x)
Sqrt (x)

Opérations au niveau de l’élément, binaire

ElementTimes (x, y)
x .* y
Minus (x, y)
x - y
Plus (x, y)
x + y`
LogPlus (x, y)
Less (x, y)
Equal (x, y)
Greater (x, y)
GreaterEqual (x, y)
NotEqual (x, y)
LessEqual (x, y)
BS.Boolean.And (a, b)
BS.Boolean.Or (a, b)
BS.Boolean.Xor (a, b)

Opérations au niveau des éléments, ternaires

BS.Boolean.If (condition, thenVal, elseVal)

Opérations de produit et de convolution de matrice

Times(A, B, outputRank=1)
A * B
TransposeTimes(A, B, outputRank=1)
Convolution(weights, x, kernelShape, mapDims=(0), stride=(1), sharing=(true), autoPadding=(true), lowerPadding=(0), upperPadding=(0), imageLayout='CHW', maxTempMemSizeInSamples=0)
Pooling(x, poolKind/*'max'|'average'*/, kernelShape, stride=(1), autoPadding=(true), lowerPadding=(0), upperPadding=(0), imageLayout='CHW')
ROIPooling(x, rois, roiOutputShape, spatialScale=1.0/16.0)

Paramètres et constantes appris

ParameterTensor {shape, learningRateMultiplier=1.0, init='uniform'/*|gaussian*/, initValueScale=1.0, initValue=0.0, randomSeed=-1, initFromFilePath=''}
Constant {scalarValue, rows = 1, cols = 1}
BS.Constants.Zero, BS.Constants.One
BS.Constants.True, BS.Constants.False, BS.Constants.None
BS.Constants.OnesTensor (shape)
BS.Constants.ZeroSequenceLike (x)

Entrées

Input (shape, dynamicAxis='', sparse=false, tag='feature')
DynamicAxis{}
EnvironmentInput (propertyName)
Mean (x), InvStdDev (x)

Fonctions de perte et métriques

CrossEntropyWithSoftmax (targetDistribution, nonNormalizedLogClassPosteriors)
CrossEntropy (targetDistribution, classPosteriors)
Logistic (label, probability)
WeightedLogistic(label, probability, instanceWeight)
ClassificationError (labels, nonNormalizedLogClassPosteriors)
MatrixL1Reg(matrix)
MatrixL2Reg(matrix)
SquareError (x, y)

Réductions

ReduceSum (z, axis=None)
ReduceLogSum (z, axis=None)
ReduceMean (z, axis=None)
ReduceMin (z, axis=None)
ReduceMax (z, axis=None)
CosDistance (x, y)
SumElements (z)

Opérations d’entraînement

BatchNormalization (input, scale, bias, runMean, runInvStdDev, spatial, normalizationTimeConstant = 0, blendTimeConstant = 0, epsilon = 0.00001, useCntkEngine = true, imageLayout='CHW')
Dropout (x)
Stabilize (x, enabled=true)
StabilizeElements (x, inputDim=x.dim, enabled=true)
CosDistanceWithNegativeSamples (x, y, numShifts, numNegSamples)

Remodelage des opérations

CNTK2.Reshape (x, shape, beginAxis=0, endAxis=0)
ReshapeDimension (x, axis, shape) = CNTK2.Reshape (x, shape, beginAxis=axis, endAxis=axis + 1)
FlattenDimensions (x, axis, num) = CNTK2.Reshape (x, 0, beginAxis=axis, endAxis=axis + num)
SplitDimension (x, axis, N) = ReshapeDimension (x, axis, 0:N)
Slice (beginIndex, endIndex, input, axis=1)
BS.Sequences.First (x) = Slice (0, 1, x, axis=-1)
BS.Sequences.Last (x) = Slice (-1, 0, x, axis=-1)
Splice (inputs, axis=1)
TransposeDimensions (x, axis1, axis2)
Transpose (x) = TransposeDimensions (x, 1, 2)
BS.Sequences.BroadcastSequenceAs (type, data1)
BS.Sequences.Gather (where, x)
BS.Sequences.Scatter (where, y)
BS.Sequences.IsFirst (x)
BS.Sequences.IsLast (x)

Récurrence

OptimizedRNNStack(weights, input, hiddenDims, numLayers=1, bidirectional=false, recurrentOp='lstm')
BS.Loop.Previous (x, timeStep=1, defaultHiddenActivation=0)
PastValue (shape, x, defaultHiddenActivation=0.1, ...) = BS.Loop.Previous (0, shape, ...)
BS.Loop.Next (x, timeStep=1, defaultHiddenActivation=0)
FutureValue (shape, x, defaultHiddenActivation=0.1, ...) = BS.Loop.Next (0, shape, ...)
LSTMP (outputDim, cellDim=outputDim, x, inputDim=x.shape, aux=BS.Constants.None, auxDim=aux.shape, prevState, enableSelfStabilization=false)
BS.Boolean.Toggle (clk, initialValue=BS.Constants.False)
BS.RNNs.RecurrentLSTMP (outputDim, cellDim=outputDim, x, inputDim=x.shape, previousHook=BS.RNNs.PreviousHC, augmentInputHook=NoAuxInputHook, augmentInputDim=0, layerIndex=0, enableSelfStabilization=false)
BS.RNNs.RecurrentLSTMPStack (layerShapes, cellDims=layerShapes, input, inputShape=input.shape, previousHook=PreviousHC, augmentInputHook=NoAuxInputHook, augmentInputShape=0, enableSelfStabilization=false)
BS.RNNs.RecurrentBirectionalLSTMPStack (layerShapes, cellDims=layerShapes, input, inputShape=input.dim, previousHook=PreviousHC, nextHook=NextHC, enableSelfStabilization=false)

Prise en charge de séquence à séquence

BS.Seq2Seq.CreateAugmentWithFixedWindowAttentionHook (attentionDim, attentionSpan, decoderDynamicAxis, encoderOutput, enableSelfStabilization=false)
BS.Seq2Seq.GreedySequenceDecoderFrom (modelAsTrained)
BS.Seq2Seq.BeamSearchSequenceDecoderFrom (modelAsTrained, beamDepth)

Opérations à usage spécial

ClassBasedCrossEntropyWithSoftmax (labelClassDescriptorVectorSequence, mainInputInfo, mainWeight, classLogProbsBeforeSoftmax)

Modification du modèle

BS.Network.Load (pathName)
BS.Network.Edit (inputModel, editFunctions, additionalRoots)
BS.Network.CloneFunction (inputNodes, outputNodes, parameters="learnable" /*|"constant"|"shared"*/)

Autre

Fail (what)
IsSameObject (a, b)
Trace (node, say='', logFrequency=traceFrequency, logFirst=10, logGradientToo=false, onlyUpToRow=100000000, onlyUpToT=100000000, format=[])

Obsolescent

ErrorPrediction (labels, nonNormalizedLogClassPosteriors)
ColumnElementTimes (...) = ElementTimes (...)
DiagTimes (...) = ElementTimes (...)
LearnableParameter(...) = Parameter(...)
LookupTable (embeddingMatrix, inputTensor)
RowRepeat (input, numRepeats)
RowSlice (beginIndex, numRows, input) = Slice(beginIndex, beginIndex + numRows, input, axis = 1)
RowStack (inputs)
RowElementTimes (...) = ElementTimes (...)
Scale (...) = ElementTimes (...)
ConstantTensor (scalarVal, shape)
Parameter (outputDim, inputDim, ...) = ParameterTensor ((outputDim:input), ...)
WeightParam (outputDim, inputDim) = Parameter (outputDim, inputDim, init='uniform', initValueScale=1, initOnCPUOnly=true, randomSeed=1)
DiagWeightParam (outputDim) = ParameterTensor ((outputDim), init='uniform', initValueScale=1, initOnCPUOnly=true, randomSeed=1)
BiasParam (dim) = ParameterTensor ((dim), init='fixedValue', value=0.0)
ScalarParam() = BiasParam (1)
SparseInput (shape, dynamicAxis='', tag='feature')
ImageInput (imageWidth, imageHeight, imageChannels, imageLayout='CHW', dynamicAxis='', tag='feature')
SparseImageInput (imageWidth, imageHeight, imageChannels, imageLayout='CHW', dynamicAxis='', tag='feature')
MeanVarNorm(feat) = PerDimMeanVarNormalization(feat, Mean (feat), InvStdDev (feat))
PerDimMeanVarNormalization (x, mean, invStdDev),
PerDimMeanVarDeNormalization (x, mean, invStdDev)
ReconcileDynamicAxis (dataInput, layoutInput)

Partager via