/// <summary>
/// Fill a sparse DMatrix using CSR compression.
/// See http://docs.scipy.org/doc/scipy/reference/generated/scipy.sparse.csr_matrix.html.
/// </summary>
private DMatrix FillSparseMatrix(IChannel ch, int nbDim, long nbRows, RoleMappedData data,
out Float[] labels, out uint[] groupCount)
{
// Allocation.
if ((2 * nbRows) >= Utils.ArrayMaxSize)
{
throw _host.Except("The training dataset is too big to hold in memory. " +
"2 features multiplied by the number of rows must be less than {0}.", Utils.ArrayMaxSize);
}
var features = new Float[nbRows * 2];
var indices = new uint[features.Length];
var indptr = new ulong[nbRows + 1];
long nelem = 0;
labels = new Float[nbRows];
var hasWeights = data.Schema.Weight != null;
var hasGroup = data.Schema.Group != null;
var weights = hasWeights ? new Float[nbRows] : null;
var groupsML = hasGroup ? new uint[nbRows] : null;
groupCount = hasGroup ? new uint[nbRows] : null;
var groupId = hasGroup ? new HashSet <uint>() : null;
int count = 0;
int lastGroup = -1;
var flags = CursOpt.Features | CursOpt.Label | CursOpt.AllowBadEverything | CursOpt.Weight | CursOpt.Group;
var featureVector = default(VBuffer <float>);
var labelProxy = float.NaN;
var groupProxy = ulong.MaxValue;
using (var cursor = data.CreateRowCursor(flags, null))
{
var featureGetter = cursor.GetFeatureFloatVectorGetter(data);
var labelGetter = cursor.GetLabelFloatGetter(data);
var weighGetter = cursor.GetOptWeightFloatGetter(data);
var groupGetter = cursor.GetOptGroupGetter(data);
while (cursor.MoveNext())
{
featureGetter(ref featureVector);
labelGetter(ref labelProxy);
labels[count] = labelProxy;
if (Single.IsNaN(labels[count]))
{
continue;
}
indptr[count] = (ulong)nelem;
int nbValues = featureVector.Count;
if (nbValues > 0)
{
if (nelem + nbValues > features.Length)
{
long newSize = Math.Max(nelem + nbValues, features.Length * 2);
if (newSize >= Utils.ArrayMaxSize)
{
throw _host.Except("The training dataset is too big to hold in memory. " +
"It should be half of {0}.", Utils.ArrayMaxSize);
}
Array.Resize(ref features, (int)newSize);
Array.Resize(ref indices, (int)newSize);
}
Array.Copy(featureVector.Values, 0, features, nelem, nbValues);
if (featureVector.IsDense)
{
for (int i = 0; i < nbValues; ++i)
{
indices[nelem++] = (uint)i;
}
}
else
{
for (int i = 0; i < nbValues; ++i)
{
indices[nelem++] = (uint)featureVector.Indices[i];
}
}
}
if (hasWeights)
{
weighGetter(ref weights[count]);
}
if (hasGroup)
{
groupGetter(ref groupProxy);
if (groupProxy >= uint.MaxValue)
{
throw _host.Except($"Group is above {uint.MaxValue}");
}
groupsML[count] = (uint)groupProxy;
if (count == 0 || groupsML[count - 1] != groupsML[count])
{
groupCount[++lastGroup] = 1;
ch.Check(!groupId.Contains(groupsML[count]), "Group Id are not contiguous.");
groupId.Add(groupsML[count]);
}
else
{
++groupCount[lastGroup];
}
}
++count;
}
}
indptr[count] = (uint)nelem;
if (nelem < features.Length * 3 / 4)
{
Array.Resize(ref features, (int)nelem);
Array.Resize(ref indices, (int)nelem);
}
PostProcessLabelsBeforeCreatingXGBoostContainer(ch, data, labels);
// We create a DMatrix.
DMatrix dtrain = new DMatrix((uint)nbDim, indptr, indices, features, (uint)count, (uint)nelem, labels: labels, weights: weights, groups: groupCount);
return(dtrain);
}
public void Predict(ref VBuffer <Float> features,
ref VBuffer <Float> predictedValues,
ref XGBoostTreeBuffer internalBuffer,
bool outputMargin = true,
int ntreeLimit = 0)
{
PredictOneOff(ref features, ref predictedValues, ref internalBuffer, outputMargin, ntreeLimit);
#if (DEBUG && MORE_CHECKING)
// This part checks that the function PredictOneOff which relies on a customized version
// of XGBoost produces the same result as the official API.
// This makes the prediction terribly slow as the prediction are called twice
// and the second call (PredictN) cannot be parallelized (lock protected).
VBuffer <Float> check = new VBuffer <float>();
DMatrix data;
if (features.IsDense)
{
data = new DMatrix(features.Values, 1, (uint)features.Count);
}
else
{
int nb = features.Count;
var indptr = new ulong[] { 0, (uint)nb };
var indices = new uint[nb];
for (int i = 0; i < nb; ++i)
{
indices[i] = (uint)features.Indices[i];
}
data = new DMatrix((uint)features.Length, indptr, indices, features.Values, 1, (uint)nb);
}
PredictN(data, ref check, outputMargin, ntreeLimit);
if (check.Count != predictedValues.Count)
{
string message =
string.Format(
"Count={0} Length={1} IsDense={2}\nValues={3}\nIndices={4}\nCustom Ouput={5}\nOfficial API={6}",
features.Count, features.Length, features.IsDense,
features.Values == null
? ""
: string.Join(", ", features.Values.Select(c => c.ToString()).ToArray()),
features.Indices == null
? ""
: string.Join(", ", features.Indices.Select(c => c.ToString()).ToArray()),
predictedValues.Values == null
? ""
: string.Join(", ", predictedValues.Values.Select(c => c.ToString()).ToArray()),
check.Values == null
? ""
: string.Join(", ", check.Values.Select(c => c.ToString()).ToArray()));
throw Contracts.Except("Mismatch between official API and custom API (dimension).\n" + message);
}
for (int i = 0; i < check.Count; ++i)
{
if (Math.Abs(check.Values[0] - predictedValues.Values[0]) > 1e-5)
{
string message =
string.Format(
"Count={0} Length={1} IsDense={2}\nValues={3}\nIndices={4}\nCustom Ouput={5}\nOfficial API={6}",
features.Count, features.Length, features.IsDense,
features.Values == null
? ""
: string.Join(", ", features.Values.Select(c => c.ToString()).ToArray()),
features.Indices == null
? ""
: string.Join(", ", features.Indices.Select(c => c.ToString()).ToArray()),
predictedValues.Values == null
? ""
: string.Join(", ", predictedValues.Values.Select(c => c.ToString()).ToArray()),
check.Values == null
? ""
: string.Join(", ", check.Values.Select(c => c.ToString()).ToArray()));
PredictOneOff(ref features, ref predictedValues, ref internalBuffer, outputMargin, ntreeLimit);
message += string.Format("\nSecond computation\n{0}", predictedValues.Values == null
? ""
: string.Join(", ", predictedValues.Values.Select(c => c.ToString()).ToArray()));
throw Contracts.Except("Mismatch between official API and custom API (output).\n" + message);
}
}
#endif
}
private DMatrix FillDenseMatrix(IChannel ch, int nbDim, long nbRows,
RoleMappedData data, out Float[] labels, out uint[] groupCount)
{
// Allocation.
string errorMessageGroup = string.Format("Group is above {0}.", uint.MaxValue);
if (nbDim * nbRows >= Utils.ArrayMaxSize)
{
throw _host.Except("The training dataset is too big to hold in memory. " +
"Number of features ({0}) multiplied by the number of rows ({1}) must be less than {2}.", nbDim, nbRows, Utils.ArrayMaxSize);
}
var features = new Float[nbDim * nbRows];
labels = new Float[nbRows];
var hasWeights = data.Schema.Weight != null;
var hasGroup = data.Schema.Group != null;
var weights = hasWeights ? new Float[nbRows] : null;
var groupsML = hasGroup ? new uint[nbRows] : null;
groupCount = hasGroup ? new uint[nbRows] : null;
var groupId = hasGroup ? new HashSet <uint>() : null;
int count = 0;
int lastGroup = -1;
int fcount = 0;
var flags = CursOpt.Features | CursOpt.Label | CursOpt.AllowBadEverything | CursOpt.Weight | CursOpt.Group;
var featureVector = default(VBuffer <float>);
var labelProxy = float.NaN;
var groupProxy = ulong.MaxValue;
using (var cursor = data.CreateRowCursor(flags, null))
{
var featureGetter = cursor.GetFeatureFloatVectorGetter(data);
var labelGetter = cursor.GetLabelFloatGetter(data);
var weighGetter = cursor.GetOptWeightFloatGetter(data);
var groupGetter = cursor.GetOptGroupGetter(data);
while (cursor.MoveNext())
{
featureGetter(ref featureVector);
labelGetter(ref labelProxy);
labels[count] = labelProxy;
if (Single.IsNaN(labels[count]))
{
continue;
}
featureVector.CopyTo(features, fcount, Single.NaN);
fcount += featureVector.Count;
if (hasWeights)
{
weighGetter(ref weights[count]);
}
if (hasGroup)
{
groupGetter(ref groupProxy);
_host.Check(groupProxy < uint.MaxValue, errorMessageGroup);
groupsML[count] = (uint)groupProxy;
if (count == 0 || groupsML[count - 1] != groupsML[count])
{
groupCount[++lastGroup] = 1;
ch.Check(!groupId.Contains(groupsML[count]), "Group Id are not contiguous.");
groupId.Add(groupsML[count]);
}
else
{
++groupCount[lastGroup];
}
}
++count;
}
}
PostProcessLabelsBeforeCreatingXGBoostContainer(ch, data, labels);
// We create a DMatrix.
DMatrix dtrain = new DMatrix(features, (uint)count, (uint)nbDim, labels: labels, weights: weights, groups: groupCount);
return(dtrain);
}
/// <summary>
/// Train and returns a booster.
/// </summary>
/// <param name="ch">IChannel</param>
/// <param name="pch">IProgressChannel</param>
/// <param name="numberOfTrees">Number of trained trees</param>
/// <param name="parameters">Parameters see <see cref="XGBoostArguments"/></param>
/// <param name="dtrain">Training set</param>
/// <param name="numBoostRound">Number of trees to train</param>
/// <param name="obj">Custom objective</param>
/// <param name="maximize">Whether to maximize feval.</param>
/// <param name="verboseEval">Requires at least one item in evals.
/// If "verbose_eval" is True then the evaluation metric on the validation set is
/// printed at each boosting stage.</param>
/// <param name="xgbModel">For continuous training.</param>
/// <param name="saveBinaryDMatrix">Save DMatrix in binary format (for debugging purpose).</param>
public static Booster Train(IChannel ch, IProgressChannel pch, out int numberOfTrees,
Dictionary <string, string> parameters, DMatrix dtrain, int numBoostRound = 10,
Booster.FObjType obj = null, bool maximize = false,
bool verboseEval = true, Booster xgbModel = null,
string saveBinaryDMatrix = null)
{
#if (!XGBOOST_RABIT)
if (WrappedXGBoostInterface.RabitIsDistributed() == 1)
{
var pname = WrappedXGBoostInterface.RabitGetProcessorName();
ch.Info("[WrappedXGBoostTraining.Train] start {0}:{1}", pname, WrappedXGBoostInterface.RabitGetRank());
}
#endif
if (!string.IsNullOrEmpty(saveBinaryDMatrix))
{
dtrain.SaveBinary(saveBinaryDMatrix);
}
Booster bst = new Booster(parameters, dtrain, xgbModel);
int numParallelTree = 1;
int nboost = 0;
if (parameters != null && parameters.ContainsKey("num_parallel_tree"))
{
numParallelTree = Convert.ToInt32(parameters["num_parallel_tree"]);
nboost /= numParallelTree;
}
if (parameters.ContainsKey("num_class"))
{
int numClass = Convert.ToInt32(parameters["num_class"]);
nboost /= numClass;
}
var prediction = new VBuffer <Float>();
var grad = new VBuffer <Float>();
var hess = new VBuffer <Float>();
var start = DateTime.Now;
#if (!XGBOOST_RABIT)
int version = bst.LoadRabitCheckpoint();
ch.Check(WrappedXGBoostInterface.RabitGetWorldSize() != 1 || version == 0);
#else
int version = 0;
#endif
int startIteration = version / 2;
nboost += startIteration;
int logten = 0;
int temp = numBoostRound * 5;
while (temp > 0)
{
logten += 1;
temp /= 10;
}
temp = Math.Max(logten - 2, 0);
logten = 1;
while (temp-- > 0)
{
logten *= 10;
}
var metrics = new List <string>()
{
"Iteration", "Training Time"
};
var units = new List <string>()
{
"iterations", "seconds"
};
if (verboseEval)
{
metrics.Add("Training Error");
metrics.Add(parameters["objective"]);
}
var header = new ProgressHeader(metrics.ToArray(), units.ToArray());
int iter = 0;
double trainTime = 0;
double trainError = double.NaN;
pch.SetHeader(header, e =>
{
e.SetProgress(0, iter, numBoostRound - startIteration);
e.SetProgress(1, trainTime);
if (verboseEval)
{
e.SetProgress(2, trainError);
}
});
for (iter = startIteration; iter < numBoostRound; ++iter)
{
if (version % 2 == 0)
{
bst.Update(dtrain, iter, ref grad, ref hess, ref prediction, obj);
#if (!XGBOOST_RABIT)
bst.SaveRabitCheckpoint();
#endif
version += 1;
}
#if (!XGBOOST_RABIT)
ch.Check(WrappedXGBoostInterface.RabitGetWorldSize() == 1 ||
version == WrappedXGBoostInterface.RabitVersionNumber());
#endif
nboost += 1;
trainTime = (DateTime.Now - start).TotalMilliseconds;
if (verboseEval)
{
pch.Checkpoint(new double?[] { iter, trainTime, trainError });
if (iter == startIteration || iter == numBoostRound - 1 || iter % logten == 0 ||
(DateTime.Now - start) > TimeSpan.FromMinutes(2))
{
string strainError = bst.EvalSet(new[] { dtrain }, new[] { "Train" }, iter);
// Example: "[0]\tTrain-error:0.028612"
if (!string.IsNullOrEmpty(strainError) && strainError.Contains(":"))
{
double val;
if (double.TryParse(strainError.Split(':').Last(), out val))
{
trainError = val;
}
}
}
}
else
{
pch.Checkpoint(new double?[] { iter, trainTime });
}
version += 1;
}
numberOfTrees = numBoostRound * numParallelTree;
if (WrappedXGBoostInterface.RabitIsDistributed() == 1)
{
var pname = WrappedXGBoostInterface.RabitGetProcessorName();
ch.Info("[WrappedXGBoostTraining.Train] end {0}:{1}", pname, WrappedXGBoostInterface.RabitGetRank());
}
return(bst);
}
