Add new lattice reader test
This commit is contained in:
Vadim Mazalov
Родитель
60c299647f
Коммит
d38c060f07
|
|
@ -140,7 +140,7 @@ public:
|
|||
tempmatrix = loglikelihood.ColumnSlice(ts, numframes);
|
||||
// if (m_deviceid == CPUDEVICE)
|
||||
{
|
||||
CopyFromCNTKMatrixToSSEMatrix(tempmatrix, numframes, predstripe, false);
|
||||
CopyFromCNTKMatrixToSSEMatrix(tempmatrix, numframes, predstripe);
|
||||
}
|
||||
|
||||
if (m_deviceid != CPUDEVICE)
|
||||
|
|
@ -176,7 +176,7 @@ public:
|
|||
|
||||
// if (doreferencealign || m_deviceid == CPUDEVICE)
|
||||
{
|
||||
CopyFromCNTKMatrixToSSEMatrix(tempmatrix, numframes, predstripe, false);
|
||||
CopyFromCNTKMatrixToSSEMatrix(tempmatrix, numframes, predstripe);
|
||||
}
|
||||
|
||||
if (m_deviceid != CPUDEVICE)
|
||||
|
|
|
|||
|
|
@ -0,0 +1,7 @@
|
|||
m1 = LoadModel("$currModel$", format="cntk")
|
||||
SetDefaultModel(m1)
|
||||
HDim = 512
|
||||
HL$newLayer$ = DNNLayer(HDim, HDim, HL$currLayer$.y)
|
||||
SetInput(OL.t, 1, HL$newLayer$.y)
|
||||
SetInput(HL$newLayer$.t, 1, HL$currLayer$.y)
|
||||
SaveModel(m1, "$newModel$", format="cntk")
|
||||
|
|
@ -0,0 +1,193 @@
|
|||
precision = "float"
|
||||
deviceId = $DeviceId$
|
||||
command = dptPre1:addLayer2:dptPre2:addLayer3:speechTrain:replaceCriterionNode:sequenceTrain
|
||||
|
||||
ndlMacros = "$ConfigDir$/macros.txt"
|
||||
|
||||
globalMeanPath = "GlobalStats/mean.363"
|
||||
globalInvStdPath = "GlobalStats/var.363"
|
||||
globalPriorPath = "GlobalStats/prior.132"
|
||||
|
||||
traceLevel = 1
|
||||
truncated = false
|
||||
|
||||
# Default SGD value used for pre-training.
|
||||
SGD = [
|
||||
epochSize = 81920
|
||||
minibatchSize = 256
|
||||
learningRatesPerMB = 0.8
|
||||
numMBsToShowResult = 10
|
||||
momentumPerMB = 0.9
|
||||
dropoutRate = 0.0
|
||||
maxEpochs = 2
|
||||
]
|
||||
|
||||
dptPre1 = [
|
||||
action = "train"
|
||||
modelPath = "$RunDir$/models/Pre1/cntkSpeech"
|
||||
|
||||
NDLNetworkBuilder = [
|
||||
networkDescription = "$ConfigDir$/dnn_1layer.txt"
|
||||
]
|
||||
]
|
||||
|
||||
addLayer2 = [
|
||||
action = "edit"
|
||||
currLayer = 1
|
||||
newLayer = 2
|
||||
currModel = "$RunDir$/models/Pre1/cntkSpeech"
|
||||
newModel = "$RunDir$/models/Pre2/cntkSpeech.0"
|
||||
editPath = "$ConfigDir$/add_layer.mel"
|
||||
]
|
||||
|
||||
dptPre2 = [
|
||||
action = "train"
|
||||
modelPath = "$RunDir$/models/Pre2/cntkSpeech"
|
||||
|
||||
NDLNetworkBuilder = [
|
||||
networkDescription = "$ConfigDir$/dnn_1layer.txt"
|
||||
]
|
||||
]
|
||||
|
||||
AddLayer3 = [
|
||||
action = "edit"
|
||||
currLayer = 2
|
||||
newLayer = 3
|
||||
currModel = "$RunDir$/models/Pre2/cntkSpeech"
|
||||
newModel = "$RunDir$/models/cntkSpeech.0"
|
||||
editPath = "$ConfigDir$/add_layer.mel"
|
||||
]
|
||||
|
||||
speechTrain = [
|
||||
action = "train"
|
||||
modelPath = "$RunDir$/models/cntkSpeech"
|
||||
#deviceId = $DeviceId$
|
||||
traceLevel = 1
|
||||
|
||||
NDLNetworkBuilder = [
|
||||
networkDescription = "$ConfigDir$/dnn.txt"
|
||||
]
|
||||
|
||||
SGD = [
|
||||
epochSize = 81920
|
||||
minibatchSize = 256:512
|
||||
learningRatesPerMB = 0.8:1.6
|
||||
numMBsToShowResult = 10
|
||||
momentumPerSample = 0.999589
|
||||
dropoutRate = 0.0
|
||||
maxEpochs = 4
|
||||
|
||||
gradUpdateType = "none"
|
||||
normWithAveMultiplier = true
|
||||
clippingThresholdPerSample = 1#INF
|
||||
]
|
||||
]
|
||||
|
||||
reader = [
|
||||
readerType = "HTKMLFReader"
|
||||
readMethod = "blockRandomize"
|
||||
miniBatchMode = "partial"
|
||||
randomize = "auto"
|
||||
verbosity = 0
|
||||
|
||||
features = [
|
||||
dim = 363
|
||||
type = "real"
|
||||
scpFile = "$DataDir$/glob_0000.scp"
|
||||
]
|
||||
|
||||
labels = [
|
||||
mlfFile = "$DataDir$/glob_0000.mlf"
|
||||
labelMappingFile = "$DataDir$/state.list"
|
||||
|
||||
labelDim = 132
|
||||
labelType = "category"
|
||||
]
|
||||
]
|
||||
|
||||
sequenceTrain = [
|
||||
action = "train"
|
||||
modelPath = $RunDir$/models/cntkSpeech.sequence
|
||||
traceLevel = 1
|
||||
|
||||
SGD = [
|
||||
epochSize = 6000000
|
||||
minibatchSize = 800000
|
||||
learningRatesPerSample = 0.000002
|
||||
momentumPerSample = 0.999589
|
||||
dropoutRate = 0.0
|
||||
maxEpochs = 3
|
||||
gradientClippingWithTruncation = true
|
||||
clippingThresholdPerSample = 1.0
|
||||
]
|
||||
reader = [
|
||||
verbosity = 0
|
||||
randomize = false
|
||||
# A list of deserializers the reader uses.
|
||||
deserializers = (
|
||||
[
|
||||
type = "HTKFeatureDeserializer"
|
||||
module = "HTKDeserializers"
|
||||
input = [
|
||||
# Description of input stream to feed the Input node named "features"
|
||||
features = [
|
||||
dim=363
|
||||
scpFile = "$DataDir$/glob_0000.scp"
|
||||
]
|
||||
]
|
||||
]:
|
||||
[
|
||||
type = "HTKMLFDeserializer"
|
||||
module = "HTKDeserializers"
|
||||
input = [
|
||||
# Description of input stream to feed the Input node named "labels"
|
||||
labels = [
|
||||
dim = 132
|
||||
mlfFile="$DataDir$/glob_0000.mlf"
|
||||
labelMappingFile = "$DataDir$/state.list"
|
||||
]
|
||||
]
|
||||
]:
|
||||
[
|
||||
type = "LatticeDeserializer"
|
||||
module = "HTKDeserializers"
|
||||
input = [
|
||||
lattice=[
|
||||
latticeIndexFile="$DataDir$/latticeIndex.txt"
|
||||
]
|
||||
]
|
||||
]
|
||||
)
|
||||
]
|
||||
|
||||
|
||||
BrainScriptNetworkBuilder = {
|
||||
baseFeatDim = 33
|
||||
featDim = 11 * baseFeatDim
|
||||
labelDim = 132
|
||||
|
||||
latticeAxis = DynamicAxis()
|
||||
features = Input{featDim}
|
||||
labels = Input{labelDim}
|
||||
lattice = Input{1,dynamicAxis=latticeAxis}
|
||||
|
||||
featExtNetwork = BS.Network.Load("$RunDir$/models/cntkSpeech")
|
||||
featExt = BS.Network.CloneFunction (
|
||||
(featExtNetwork.features),
|
||||
[netEval = featExtNetwork.OL_z;scaledLogLikelihood = featExtNetwork.scaledLogLikelihood ],
|
||||
parameters="learnable")
|
||||
|
||||
clonedmodel= featExt(features)
|
||||
|
||||
cr = SequenceWithLattice(labels, clonedmodel.netEval, clonedmodel.scaledLogLikelihood, lattice, "$DataDir$/CY2SCH010061231_1369712653.numden.lats.symlist", "$DataDir$/model.overalltying", "$DataDir$/state.list", "$DataDir$/model.transprob", tag="criterion")
|
||||
|
||||
Err = ClassificationError(labels,clonedmodel.netEval,tag="evaluation");
|
||||
|
||||
}
|
||||
|
||||
]
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
|
@ -0,0 +1,42 @@
|
|||
load = ndlMacroDefine
|
||||
run = DNN
|
||||
|
||||
ndlMacroDefine = [
|
||||
# Macro definitions
|
||||
MeanVarNorm(x) = [
|
||||
xMean = Mean(x);
|
||||
xStdDev = InvStdDev(x)
|
||||
xNorm = PerDimMeanVarNormalization(x,xMean,xStdDev)
|
||||
]
|
||||
]
|
||||
|
||||
DNN = [
|
||||
#define basic i/o
|
||||
featDim = 363
|
||||
labelDim = 132
|
||||
hiddenDim = 512
|
||||
|
||||
features = Input(featDim, tag="feature")
|
||||
labels = Input(labelDim, tag="label")
|
||||
|
||||
globalMean = Parameter(featDim, 1, init="fromFile", initFromFilePath="$globalMeanPath$", computeGradient=false)
|
||||
globalInvStd = Parameter(featDim, 1, init="fromFile", initFromFilePath="$globalInvStdPath$", computeGradient=false)
|
||||
globalPrior = Parameter(labelDim, 1, init="fromFile", initFromFilePath="$globalPriorPath$", computeGradient=false)
|
||||
logPrior = Log(globalPrior)
|
||||
|
||||
# define network
|
||||
featNorm = PerDimMeanVarNormalization(features, globalMean, globalInvStd)
|
||||
|
||||
# layer 1 363 X 512
|
||||
z1 = DNNLayer(featDim, hiddenDim, featNorm);
|
||||
# layer 2 512 X 512
|
||||
z2 = DNNLayer(hiddenDim, hiddenDim, z1);
|
||||
# layer 3 512 X 512
|
||||
z3 = DNNLayer(hiddenDim, hiddenDim, z2);
|
||||
# last layer 512 X 132
|
||||
z4 = DNNLastLayer(hiddenDim, labelDim, z3);
|
||||
|
||||
ce = CrossEntropyWithSoftmax(labels, z4, tag="criterion");
|
||||
err = ClassificationError(labels, z4, tag="evaluation");
|
||||
scaledLogLikelihood = Minus(z4, logPrior, tag="output")
|
||||
]
|
||||
|
|
@ -0,0 +1,38 @@
|
|||
load = ndlMacroDefine
|
||||
run = DNN
|
||||
|
||||
ndlMacroDefine = [
|
||||
# Macro definitions
|
||||
MeanVarNorm(x) = [
|
||||
xMean = Mean(x);
|
||||
xStdDev = InvStdDev(x)
|
||||
xNorm = PerDimMeanVarNormalization(x, xMean, xStdDev)
|
||||
]
|
||||
]
|
||||
|
||||
DNN = [
|
||||
#define basic i/o
|
||||
featDim = 363
|
||||
labelDim = 132
|
||||
hiddenDim = 512
|
||||
|
||||
features = Input(featDim, tag="feature")
|
||||
labels = Input(labelDim, tag="label")
|
||||
|
||||
globalMean = Parameter(featDim, 1, init="fromFile", initFromFilePath="$globalMeanPath$", computeGradient=false)
|
||||
globalInvStd = Parameter(featDim, 1, init="fromFile", initFromFilePath="$globalInvStdPath$", computeGradient=false)
|
||||
globalPrior = Parameter(labelDim, 1, init="fromFile", initFromFilePath="$globalPriorPath$", computeGradient=false)
|
||||
logPrior = Log(globalPrior)
|
||||
|
||||
# define network
|
||||
featNorm = PerDimMeanVarNormalization(features, globalMean, globalInvStd)
|
||||
|
||||
# layer 1 363 X 512
|
||||
HL1 = DNNLayer(featDim, hiddenDim, featNorm);
|
||||
# last layer 512 X 132
|
||||
OL = DNNLastLayer(hiddenDim, labelDim, HL1);
|
||||
|
||||
ce = CrossEntropyWithSoftmax(labels, OL, tag="criterion");
|
||||
err = ClassificationError(labels, OL, tag="evaluation");
|
||||
scaledLogLikelihood = Minus(OL, logPrior, tag="output")
|
||||
]
|
||||
|
|
@ -0,0 +1,18 @@
|
|||
DNNLayer(inDim, outDim, x) = [
|
||||
#W = Parameter(outDim, inDim, init="uniform", initValueScale=1, initOnCPUOnly=true, randomSeed=1); # randomizing on CPU with fixed seed to get reproducable results across configurations
|
||||
#b = Parameter(outDim, 1, init="uniform", initValueScale=1, initOnCPUOnly=true, randomSeed=1);
|
||||
W = Parameter(outDim, inDim);
|
||||
b = Parameter(outDim, 1);
|
||||
t = Times(W, x);
|
||||
z = Plus(t, b);
|
||||
y = sigmoid(z);
|
||||
]
|
||||
|
||||
DNNLastLayer(hiddenDim, LabelDim, x) = [
|
||||
#W = Parameter(LabelDim, hiddenDim, init="uniform", initValueScale=1, initOnCPUOnly=true, randomSeed=1);
|
||||
#b = Parameter(LabelDim, 1, init="uniform", initValueScale=1, initOnCPUOnly=true, randomSeed=1);
|
||||
W = Parameter(LabelDim, hiddenDim);
|
||||
b = Parameter(LabelDim, 1);
|
||||
t = Times(W, x);
|
||||
z = Plus(t, b);
|
||||
]
|
||||
|
|
@ -0,0 +1,11 @@
|
|||
m1 = LoadModel("$currModel$", format="cntk")
|
||||
SetDefaultModel(m1)
|
||||
SetProperty(ce, "criterion", false)
|
||||
Remove(ce)
|
||||
|
||||
# note: we use the same name, since Jenkins does not allow to specify two different node names
|
||||
ce = SequenceWithSoftmax(labels, OL.z, scaledLogLikelihood)
|
||||
SetProperty(ce, criterion, true)
|
||||
|
||||
SaveModel(m1, "$newModel$", format="cntk")
|
||||
Dump(m1, "$newModel$.dump.txt")
|
||||
|
|
@ -0,0 +1,31 @@
|
|||
#!/bin/bash
|
||||
|
||||
. $TEST_ROOT_DIR/run-test-common
|
||||
|
||||
# This test uses a large dataset which is not part of the CNTK repository itself
|
||||
# We use the dataset from an external location specified using an environment variable
|
||||
if [[ "$CNTK_EXTERNAL_TESTDATA_SOURCE_DIRECTORY" == "" || ! -d "$CNTK_EXTERNAL_TESTDATA_SOURCE_DIRECTORY" ]]; then
|
||||
echo 'This test uses external data that is not part of the CNTK repository. Environment variable CNTK_EXTERNAL_TESTDATA_SOURCE_DIRECTORY must be set to point to the external test data location'
|
||||
exit 1
|
||||
fi
|
||||
|
||||
if [ "$OS" == "Windows_NT" ]; then
|
||||
DataSourceDir=`cygpath -au $CNTK_EXTERNAL_TESTDATA_SOURCE_DIRECTORY`/Speech/AN4Corpus/v0
|
||||
else
|
||||
DataSourceDir=$CNTK_EXTERNAL_TESTDATA_SOURCE_DIRECTORY/Speech/AN4Corpus/v0
|
||||
fi
|
||||
|
||||
# Copy the test data to the test run directory
|
||||
DataDir=$TEST_RUN_DIR/TestData
|
||||
mkdir $DataDir
|
||||
cp -R $DataSourceDir/* $DataDir || exit $?
|
||||
|
||||
# cntkrun <CNTK config file name> <additional CNTK args>
|
||||
cntkrun cntk_sequence.cntk
|
||||
ExitCode=$?
|
||||
|
||||
# Delete the test data
|
||||
rm -rf $DataDir
|
||||
|
||||
exit $ExitCode
|
||||
|
||||
|
|
@ -0,0 +1,31 @@
|
|||
dataDir: ../../Data
|
||||
tags:
|
||||
# Note: Sequence training is currently not supported on the CPU
|
||||
- bvt-s (build_sku == 'gpu') and (device == 'gpu') and ((flavor == 'debug') ^ (os == 'windows'))
|
||||
- nightly-s (build_sku == 'gpu') and (device == 'gpu')
|
||||
- weekly-s (build_sku == 'gpu') and (device == 'gpu')
|
||||
|
||||
testCases:
|
||||
CNTK Run must be completed:
|
||||
patterns:
|
||||
- __COMPLETED__
|
||||
|
||||
Must train epochs in exactly same order and parameters:
|
||||
patterns:
|
||||
- Starting Epoch {{integer}}
|
||||
- learning rate per sample = {{float}}
|
||||
- momentum = {{float}}
|
||||
|
||||
Epochs must be finished with expected results:
|
||||
patterns:
|
||||
- Finished Epoch[{{integer}} of {{integer}}]
|
||||
- ce = {{float,tolerance=.1%}}
|
||||
- err = {{float,tolerance=.25%}}
|
||||
- learningRatePerSample = {{float,tolerance=0.001%}}
|
||||
|
||||
Per-minibatch training results must match:
|
||||
patterns:
|
||||
- Epoch[{{integer}} of {{integer}}]-Minibatch[{{integer}}-{{integer}}
|
||||
- " * {{integer}}; "
|
||||
- ce = {{float,tolerance=.2%}}
|
||||
|
||||
Загрузка…
Ссылка в новой задаче