/// <summary>
/// This method ensures that the data meets the requirements of this trainer and its
/// subclasses, injects necessary transforms, and throws if it couldn't meet them.
/// </summary>
/// <param name="ch">The channel</param>
/// <param name="examples">The training examples</param>
/// <param name="weightSetCount">Gets the length of weights and bias array. For binary classification and regression,
/// this is 1. For multi-class classification, this equals the number of classes on the label.</param>
/// <returns>A potentially modified version of <paramref name="examples"/></returns>
protected RoleMappedData PrepareDataFromTrainingExamples(IChannel ch, RoleMappedData examples, out int weightSetCount)
{
ch.AssertValue(examples);
CheckLabel(examples, out weightSetCount);
examples.CheckFeatureFloatVector();
var idvToShuffle = examples.Data;
IDataView idvToFeedTrain;
if (idvToShuffle.CanShuffle)
{
idvToFeedTrain = idvToShuffle;
}
else
{
var shuffleArgs = new RowShufflingTransformer.Arguments
{
PoolOnly = false,
ForceShuffle = ShuffleData
};
idvToFeedTrain = new RowShufflingTransformer(Host, shuffleArgs, idvToShuffle);
}
ch.Assert(idvToFeedTrain.CanShuffle);
var roles = examples.Schema.GetColumnRoleNames();
var examplesToFeedTrain = new RoleMappedData(idvToFeedTrain, roles);
ch.AssertValue(examplesToFeedTrain.Schema.Label);
ch.AssertValue(examplesToFeedTrain.Schema.Feature);
if (examples.Schema.Weight != null)
{
ch.AssertValue(examplesToFeedTrain.Schema.Weight);
}
int numFeatures = examplesToFeedTrain.Schema.Feature.Type.VectorSize;
ch.Check(numFeatures > 0, "Training set has no features, aborting training.");
return(examplesToFeedTrain);
}
IDataTransform AppendToPipeline(IDataView input)
{
IDataView current = input;
if (_shuffleInput)
{
var args1 = new RowShufflingTransformer.Arguments()
{
ForceShuffle = false,
ForceShuffleSeed = _seedShuffle,
PoolRows = _poolRows,
PoolOnly = false,
};
current = new RowShufflingTransformer(Host, args1, current);
}
// We generate a random number.
var columnName = current.Schema.GetTempColumnName();
var args2 = new GenerateNumberTransform.Arguments()
{
Column = new GenerateNumberTransform.Column[] { new GenerateNumberTransform.Column()
{
Name = columnName
} },
Seed = _seed ?? 42
};
IDataTransform currentTr = new GenerateNumberTransform(Host, args2, current);
// We convert this random number into a part.
var cRatios = new float[_ratios.Length];
cRatios[0] = 0;
for (int i = 1; i < _ratios.Length; ++i)
{
cRatios[i] = cRatios[i - 1] + _ratios[i - 1];
}
ValueMapper <float, int> mapper = (in float src, ref int dst) =>
{
for (int i = cRatios.Length - 1; i > 0; --i)
{
if (src >= cRatios[i])
{
dst = i;
return;
}
}
dst = 0;
};
// Get location of columnName
int index;
currentTr.Schema.TryGetColumnIndex(columnName, out index);
var ct = currentTr.Schema.GetColumnType(index);
var view = LambdaColumnMapper.Create(Host, "Key to part mapper", currentTr,
columnName, _newColumn, ct, NumberType.I4, mapper);
// We cache the result to avoid the pipeline to change the random number.
var args3 = new ExtendedCacheTransform.Arguments()
{
inDataFrame = string.IsNullOrEmpty(_cacheFile),
numTheads = _numThreads,
cacheFile = _cacheFile,
reuse = _reuse,
};
currentTr = new ExtendedCacheTransform(Host, args3, view);
// Removing the temporary column.
var finalTr = ColumnSelectingTransformer.CreateDrop(Host, currentTr, new string[] { columnName });
var taggedViews = new List <Tuple <string, ITaggedDataView> >();
// filenames
if (_filenames != null || _tags != null)
{
int nbf = _filenames == null ? 0 : _filenames.Length;
if (nbf > 0 && nbf != _ratios.Length)
{
throw Host.Except("Differen number of filenames and ratios.");
}
int nbt = _tags == null ? 0 : _tags.Length;
if (nbt > 0 && nbt != _ratios.Length)
{
throw Host.Except("Differen number of filenames and ratios.");
}
int nb = Math.Max(nbf, nbt);
using (var ch = Host.Start("Split the datasets and stores each part."))
{
for (int i = 0; i < nb; ++i)
{
if (_filenames == null || !_filenames.Any())
{
ch.Info("Create part {0}: {1} (tag: {2})", i + 1, _ratios[i], _tags[i]);
}
else
{
ch.Info("Create part {0}: {1} (file: {2})", i + 1, _ratios[i], _filenames[i]);
}
var ar1 = new RangeFilter.Arguments()
{
Column = _newColumn, Min = i, Max = i, IncludeMax = true
};
int pardId = i;
var filtView = LambdaFilter.Create <int>(Host, string.Format("Select part {0}", i), currentTr,
_newColumn, NumberType.I4,
(in int part) => { return(part.Equals(pardId)); });