private void SaveAsOnnxCore(OnnxContext ctx, string srcVariableName, string dstVariableName)
{
const int minimumOpSetVersion = 10;
ctx.CheckOpSetVersion(minimumOpSetVersion, LoaderSignature);
// StringNormalizer only takes input of shapes [C] or [1,C],
// so the input is squeezed to support inferred shapes ( e.g. [-1,C] ).
var opType = "Squeeze";
var squeezeOutput = ctx.AddIntermediateVariable(null, "SqueezeOutput", true);
var node = ctx.CreateNode(opType, srcVariableName, squeezeOutput, ctx.GetNodeName(opType), "");
node.AddAttribute("axes", new long[] { 1 });
opType = "StringNormalizer";
var normalizerOutput = ctx.AddIntermediateVariable(null, "NormalizerOutput", true);
node = ctx.CreateNode(opType, squeezeOutput, normalizerOutput, ctx.GetNodeName(opType), "");
var isCaseChange = (_parent._caseMode == TextNormalizingEstimator.CaseMode.Lower) ? "LOWER" :
(_parent._caseMode == TextNormalizingEstimator.CaseMode.Upper) ? "UPPER" : "NONE";
node.AddAttribute("case_change_action", isCaseChange);
opType = "Unsqueeze";
node = ctx.CreateNode(opType, normalizerOutput, dstVariableName, ctx.GetNodeName(opType), "");
node.AddAttribute("axes", new long[] { 1 });
}
private void SaveAsOnnxCore(OnnxContext ctx, int iinfo, string srcVariableName, string dstVariableName)
{
const int minimumOpSetVersion = 9;
ctx.CheckOpSetVersion(minimumOpSetVersion, LoaderSignature);
var model = _parent._models[iinfo];
int dimension = _srcTypes[iinfo].GetValueCount();
Host.Assert(model.Length == dimension * dimension);
var parameters = _parent._columns[iinfo];
Host.Assert(parameters.Kind == WhiteningKind.PrincipalComponentAnalysis || parameters.Kind == WhiteningKind.ZeroPhaseComponentAnalysis);
int rank = (parameters.Kind == WhiteningKind.PrincipalComponentAnalysis && parameters.Rank > 0) ? parameters.Rank : dimension;
Host.CheckParam(rank <= dimension, nameof(rank), "Rank must be at most the dimension of untransformed data.");
long[] modelDimension = { rank, dimension };
var opType = "Gemm";
var modelName = ctx.AddInitializer(model.Take(rank * dimension), modelDimension, "model");
var zeroValueName = ctx.AddInitializer((float)0);
var gemmOutput = ctx.AddIntermediateVariable(null, "GemmOutput", true);
var node = ctx.CreateNode(opType, new[] { modelName, srcVariableName, zeroValueName }, new[] { gemmOutput }, ctx.GetNodeName(opType), "");
node.AddAttribute("transB", 1);
opType = "Transpose";
ctx.CreateNode(opType, new[] { gemmOutput }, new[] { dstVariableName }, ctx.GetNodeName(opType), "");
}
private void SaveAsOnnxCore(OnnxContext ctx, int iinfo, string srcVariableName, string dstVariableName)
{
const int minimumOpSetVersion = 9;
ctx.CheckOpSetVersion(minimumOpSetVersion, LoaderSignature);
string opType = "Tokenizer";
DataViewType dataViewType;
if (_isSourceVector[iinfo])
{
dataViewType = new VectorDataViewType(TextDataViewType.Instance, _sourceVectorLength[iinfo]);
}
else
{
dataViewType = TextDataViewType.Instance;
}
string tokenizerOutput = ctx.AddIntermediateVariable(dataViewType, "TokenizerOutput", true);
var node = ctx.CreateNode(opType, srcVariableName, tokenizerOutput, ctx.GetNodeName(opType), "com.microsoft");
node.AddAttribute("mark", _parent._useMarkerChars);
node.AddAttribute("mincharnum", 1);
node.AddAttribute("pad_value", "");
node.AddAttribute("separators", new string[] { "" });
opType = "Squeeze";
var squeezeOutput = ctx.AddIntermediateVariable(dataViewType, "SqueezeOutput");
node = ctx.CreateNode(opType, tokenizerOutput, squeezeOutput, ctx.GetNodeName(opType), "");
node.AddAttribute("axes", new long[] { 1 });
opType = "LabelEncoder";
var labelEncoderOutput = ctx.AddIntermediateVariable(NumberDataViewType.Int64, "LabelEncoderOutput");
node = ctx.CreateNode(opType, squeezeOutput, labelEncoderOutput, ctx.GetNodeName(opType));
IEnumerable <string> charStrings = Enumerable.Range(0, 65535).Select(x => ((char)x).ToString());
IEnumerable <long> charValues = Enumerable.Range(0, 65535).Select(x => Convert.ToInt64(x));
node.AddAttribute("keys_strings", charStrings);
node.AddAttribute("values_int64s", charValues);
opType = "Cast";
var castNode = ctx.CreateNode(opType, labelEncoderOutput, dstVariableName, ctx.GetNodeName(opType), "");
var t = InternalDataKindExtensions.ToInternalDataKind(DataKind.UInt16).ToType();
castNode.AddAttribute("to", t);
}
bool ISingleCanSaveOnnx.SaveAsOnnx(OnnxContext ctx, string[] outputNames, string featureColumn)
{
const int minimumOpSetVersion = 9;
ctx.CheckOpSetVersion(minimumOpSetVersion, LoaderSignature);
// Mapping score to prediction
var fastTreeOutput = ctx.AddIntermediateVariable(null, "FastTreeOutput", true);
base.SaveAsOnnx(ctx, new[] { fastTreeOutput }, featureColumn);
var opType = "Exp";
ctx.CreateNode(opType, new[] { fastTreeOutput }, outputNames, ctx.GetNodeName(opType), "");
return(true);
}
public void SaveAsOnnx(OnnxContext ctx)
{
const int minimumOpSetVersion = 9;
ctx.CheckOpSetVersion(minimumOpSetVersion, LoaderSignature);
var outputToInputMap = _mapper.OutputToInputMap;
for (int i = 0; i < outputToInputMap.Length; i++)
{
var srcCol = InputSchema[outputToInputMap[i]];
var dstCol = OutputSchema[i];
if (!ctx.ContainsColumn(srcCol.Name) || dstCol.IsHidden)
{
continue;
}
var srcVariable = ctx.GetVariableName(srcCol.Name);
var dstVariable = ctx.AddIntermediateVariable(dstCol.Type, dstCol.Name);
string opType = "Identity";
ctx.CreateNode(opType, srcVariable, dstVariable, ctx.GetNodeName(opType), "");
}
}
private bool SaveAsOnnxCore(OnnxContext ctx, string srcVariableName, DataViewType columnType)
{
const int minimumOpSetVersion = 9;
ctx.CheckOpSetVersion(minimumOpSetVersion, LoaderSignature);
Type type = columnType.RawType;
int size;
if (columnType is VectorDataViewType && columnType.IsKnownSizeVector())
{
size = columnType.GetVectorSize();
}
else
{
size = 1;
}
if ((type == typeof(int)) ||
(type == typeof(short)) || (type == typeof(ushort)) ||
(type == typeof(sbyte)) || (type == typeof(byte)))
{
ctx.AddInitializer(new int[size], type, new long[] { 1, size }, srcVariableName, false);
}
else if (type == typeof(uint) || (type == typeof(ulong)))
{
ctx.AddInitializer(new ulong[size], type == typeof(ulong), new long[] { 1, size }, srcVariableName, false);
}
else if (type == typeof(bool))
{
ctx.AddInitializer(new bool[size], new long[] { 1, size }, srcVariableName, false);
}
else if (type == typeof(long))
{
ctx.AddInitializer(new long[size], new long[] { 1, size }, srcVariableName, false);
}
else if (type == typeof(float))
{
ctx.AddInitializer(new float[size], new long[] { 1, size }, srcVariableName, false);
}
else if (type == typeof(double))
{
ctx.AddInitializer(new double[size], new long[] { 1, size }, srcVariableName, false);
}
else if ((type == typeof(string)) || (columnType is TextDataViewType))
{
string[] values = new string[size];
for (int i = 0; i < size; i++)
{
values[i] = "";
}
ctx.AddInitializer(values, new long[] { 1, size }, srcVariableName, false);
}
else
{
return(false);
}
return(true);
}
/// <summary>
/// Creates an Onnx inferencing model by vectorizing and following the logic found in <see cref="Map"/>
/// </summary>
bool ISingleCanSaveOnnx.SaveAsOnnx(OnnxContext ctx, string[] outputNames, string featureColumn)
{
const int minimumOpSetVersion = 9;
ctx.CheckOpSetVersion(minimumOpSetVersion, "MulticlassNaiveBayes");
float[] featureHistogram = new float[_featureHistogram[0].Length * _labelHistogram.Length];
float[] labelHistogramExpanded = new float[_featureHistogram[0].Length * _labelHistogram.Length];
for (int i = 0; i < _featureHistogram.Length; i++)
{
Array.Copy(_featureHistogram[i], 0, featureHistogram, i * _featureHistogram[i].Length, _featureHistogram[i].Length);
}
for (int i = 0; i < _featureHistogram[0].Length; i++)
{
Array.Copy(_labelHistogram, 0, labelHistogramExpanded, i * _featureHistogram.Length, _featureHistogram.Length);
}
var one = ctx.AddInitializer(1.0f, "one");
var oneInt = ctx.AddInitializer(1, typeof(int), "oneInt");
var zero = ctx.AddInitializer(0.0f, "zero");
var labelCount = ctx.AddInitializer((float)_labelCount, "labelCount");
var trainingCount = ctx.AddInitializer((float)_totalTrainingCount, "totalTrainingCount");
var labelHistogram = ctx.AddInitializer(labelHistogramExpanded.Take(_labelHistogram.Length), new long[] { _labelHistogram.Length, 1 }, "labelHistogram");
var featureHistogramName = ctx.AddInitializer(featureHistogram, new long[] { _featureHistogram.Length, _featureHistogram[0].Length }, "featureHistogram");
var labelHistogramName = ctx.AddInitializer(labelHistogramExpanded, new long[] { _featureHistogram[0].Length, _labelHistogram.Length }, "labelHistogramExpanded");
var learnedAbsentFeatureLogProb = ctx.AddInitializer(_absentFeaturesLogProb, new long[] { _absentFeaturesLogProb.Length, 1 }, "absentFeaturesLogProb");
var typeOne = new VectorDataViewType(NumberDataViewType.Single, 1);
var typeFea = new VectorDataViewType(NumberDataViewType.Single, _featureHistogram[0].Length);
var typeLabelByFea = new VectorDataViewType(NumberDataViewType.Single, _labelHistogram.Length, _featureHistogram[0].Length);
var typeLabelByOne = new VectorDataViewType(NumberDataViewType.Single, _labelHistogram.Length, 1);
var greaterOutput = ctx.AddIntermediateVariable(new VectorDataViewType(BooleanDataViewType.Instance, _featureHistogram[0].Length), "greaterOutput");
var opType = "Greater";
ctx.CreateNode(opType, new[] { featureColumn, zero }, new[] { greaterOutput }, ctx.GetNodeName(opType), "");
opType = "Cast";
var castOutput = ctx.AddIntermediateVariable(typeFea, "CastOutput");
var node = ctx.CreateNode(opType, greaterOutput, castOutput, ctx.GetNodeName(opType), "");
var t = InternalDataKindExtensions.ToInternalDataKind(DataKind.Single).ToType();
node.AddAttribute("to", t);
opType = "ExpandDims";
var isFeaturePresent = ctx.AddIntermediateVariable(new VectorDataViewType(NumberDataViewType.Single, 1, _featureHistogram[0].Length), "isFeaturePresent");
ctx.CreateNode(opType, new[] { castOutput, oneInt }, new[] { isFeaturePresent }, ctx.GetNodeName(opType), "com.microsoft");
//initialize logProb
opType = "Div";
var divOutput = ctx.AddIntermediateVariable(typeOne, "DivOutput");
ctx.CreateNode(opType, new[] { labelHistogram, trainingCount }, new[] { divOutput }, ctx.GetNodeName(opType), "");
opType = "Log";
var logOutput = ctx.AddIntermediateVariable(typeOne, "LogOutput");
ctx.CreateNode(opType, divOutput, logOutput, ctx.GetNodeName(opType), "");
//log1
opType = "Sum";
var sumOutput = ctx.AddIntermediateVariable(_inputType, "SumOutput");
ctx.CreateNode(opType, new[] { featureHistogramName, one }, new[] { sumOutput }, ctx.GetNodeName(opType), "");
var logOutput1 = ctx.AddIntermediateVariable(typeLabelByFea, "LogOutput");
LogMul(ctx, sumOutput, isFeaturePresent, logOutput1);
//log2
opType = "Transpose";
var labelHistogramTrans = ctx.AddIntermediateVariable(typeFea, "Transpose");
ctx.CreateNode(opType, labelHistogramName, labelHistogramTrans, ctx.GetNodeName(opType), "");
opType = "Sub";
var absentFeatureCount = ctx.AddIntermediateVariable(typeFea, "AbsentFeatureCounts");
ctx.CreateNode(opType, new[] { labelHistogramTrans, featureHistogramName }, new[] { absentFeatureCount }, ctx.GetNodeName(opType), "");
opType = "Sum";
sumOutput = ctx.AddIntermediateVariable(typeFea, "SumOutput");
ctx.CreateNode(opType, new[] { labelHistogramTrans, labelCount }, new[] { sumOutput }, ctx.GetNodeName(opType), "");
var logOutput2 = ctx.AddIntermediateVariable(typeLabelByFea, "LogOutput");
LogMul(ctx, sumOutput, isFeaturePresent, logOutput2);
//log3
opType = "Sum";
sumOutput = ctx.AddIntermediateVariable(typeFea, "SumOutput");
ctx.CreateNode(opType, new[] { absentFeatureCount, one }, new[] { sumOutput }, ctx.GetNodeName(opType), "");
var logOutput3 = ctx.AddIntermediateVariable(typeLabelByFea, "LogOutput");
LogMul(ctx, sumOutput, isFeaturePresent, logOutput3);
//result
opType = "Sub";
var logProb = ctx.AddIntermediateVariable(typeLabelByFea, "LogProb");
ctx.CreateNode(opType, new[] { logOutput1, logOutput2 }, new[] { logProb }, ctx.GetNodeName(opType), "");
opType = "Sub";
var absentFeatureLogProb = ctx.AddIntermediateVariable(typeLabelByFea, "AbsentFeatureLogProb");
ctx.CreateNode(opType, new[] { logOutput3, logOutput2 }, new[] { absentFeatureLogProb }, ctx.GetNodeName(opType), "");
opType = "ReduceSum";
var logProbReduceSum = ctx.AddIntermediateVariable(typeLabelByOne, "ReduceSum");
node = ctx.CreateNode(opType, new[] { logProb }, new[] { logProbReduceSum }, ctx.GetNodeName(opType), "");
long[] list = { 2 };
node.AddAttribute("axes", list);
opType = "ReduceSum";
var absentFeatureLogProbReduceSum = ctx.AddIntermediateVariable(typeLabelByOne, "ReduceSum");
node = ctx.CreateNode(opType, new[] { absentFeatureLogProb }, new[] { absentFeatureLogProbReduceSum }, ctx.GetNodeName(opType), "");
node.AddAttribute("axes", list);
opType = "Cast";
castOutput = ctx.AddIntermediateVariable(NumberDataViewType.Single, "CastOutput");
node = ctx.CreateNode(opType, learnedAbsentFeatureLogProb, castOutput, ctx.GetNodeName(opType), "");
t = InternalDataKindExtensions.ToInternalDataKind(DataKind.Single).ToType();
node.AddAttribute("to", t);
opType = "Sub";
var subOutput = ctx.AddIntermediateVariable(typeLabelByOne, "SubOutput");
ctx.CreateNode(opType, new[] { castOutput, absentFeatureLogProbReduceSum }, new[] { subOutput }, ctx.GetNodeName(opType), "");
opType = "Sum";
sumOutput = ctx.AddIntermediateVariable(typeLabelByOne, "SumOutput");
ctx.CreateNode(opType, new[] { subOutput, logProbReduceSum, logOutput }, new[] { sumOutput }, ctx.GetNodeName(opType), "");
opType = "Squeeze";
var squeezeNode = ctx.CreateNode(opType, sumOutput, outputNames[1], ctx.GetNodeName(opType), "");
squeezeNode.AddAttribute("axes", new long[] { 2 });
opType = "ArgMax";
var scoreIndex = ctx.AddIntermediateVariable(new VectorDataViewType(NumberDataViewType.Int64, 1), "ScoreIndex");
node = ctx.CreateNode(opType, new[] { sumOutput }, new[] { scoreIndex }, ctx.GetNodeName(opType), "");
node.AddAttribute("axis", 1);
node.AddAttribute("keepdims", 0);
opType = "Cast";
castOutput = ctx.AddIntermediateVariable(typeOne, "CastOutput");
node = ctx.CreateNode(opType, scoreIndex, castOutput, ctx.GetNodeName(opType), "");
t = InternalDataKindExtensions.ToInternalDataKind(DataKind.Single).ToType();
node.AddAttribute("to", t);
//log3
opType = "Sum";
sumOutput = ctx.AddIntermediateVariable(typeOne, "SumOutput");
ctx.CreateNode(opType, new[] { castOutput, one }, new[] { sumOutput }, ctx.GetNodeName(opType), "");
opType = "Cast";
node = ctx.CreateNode(opType, sumOutput, outputNames[0], ctx.GetNodeName(opType), "");
t = InternalDataKindExtensions.ToInternalDataKind(DataKind.UInt32).ToType();
node.AddAttribute("to", t);
return(true);
}