public string ToGainSummary(FeaturesToContentMap fmap, Dictionary <int, int> featureToID, int prefix, bool includeZeroGainFeatures, bool normalize, int startingCommentNumber)
{
if (_trees.Count == 0)
{
return(string.Empty);
}
StringBuilder output = new StringBuilder();
// use only first prefix trees
if (prefix > _trees.Count || prefix < 0)
{
prefix = _trees.Count;
}
FeatureToGainMap gainMap = new FeatureToGainMap(_trees.Take(prefix).ToList(), normalize);
if (includeZeroGainFeatures)
{
for (int ifeat = 0; ifeat < fmap.Count; ++ifeat)
{
gainMap[ifeat++] += 0.0;
}
}
var sortedByGain = gainMap.OrderByDescending(pair => pair.Value).AsEnumerable();
var maxValue = sortedByGain.First().Value;
double normalizingFactor = normalize && maxValue != 0 ? Math.Sqrt(maxValue) : 1.0;
double power = normalize ? 0.5 : 1.0;
foreach (var pair in sortedByGain)
{
int outputInputId = featureToID.ContainsKey(pair.Key) ? featureToID[pair.Key] : 0;
output.Append(string.Format("C:{0}=FG:I{1}:{2}:{3}\n", startingCommentNumber++, outputInputId,
fmap.GetName(pair.Key), Math.Pow(pair.Value, power) / normalizingFactor));
}
return(output.ToString());
}
/// <summary>
/// returns the ensemble in the production TreeEnsemble format
/// </summary>
public string ToTreeEnsembleIni(FeaturesToContentMap fmap,
string trainingParams, bool appendFeatureGain, bool includeZeroGainFeatures = true)
{
StringBuilder sbEvaluator = new StringBuilder();
StringBuilder sbInput = new StringBuilder();
StringBuilder sb = new StringBuilder();
Dictionary <int, int> featureToID = new Dictionary <int, int>(); //Mapping from feature to ini input id
int numNodes = 0;
// Append the pretrained input
if (_firstInputInitializationContent != null)
{
numNodes++;
featureToID[-1] = 0;
sbInput.AppendFormat("\n[Input:1]\n{0}\n", _firstInputInitializationContent);
}
int evaluatorCounter = 0;
for (int w = 0; w < NumTrees; ++w)
{
_trees[w].ToTreeEnsembleFormat(sbEvaluator, sbInput, fmap, ref evaluatorCounter, featureToID);
}
numNodes += evaluatorCounter;
sb.AppendFormat("[TreeEnsemble]\nInputs={0}\nEvaluators={1}\n", featureToID.Count, evaluatorCounter + 1);
sb.Append(sbInput);
sb.Append(sbEvaluator);
// Append the final aggregator
sb.AppendFormat("\n[Evaluator:{0}]\nEvaluatorType=Aggregator\nNumNodes={1}\nNodes=", evaluatorCounter + 1, numNodes);
// Nodes
if (_firstInputInitializationContent != null)
{
sb.Append("I:1");
}
if (NumTrees > 0)
{
if (_firstInputInitializationContent != null)
{
sb.Append("\t");
}
sb.Append("E:1");
}
for (int w = 1; w < NumTrees; ++w)
{
sb.AppendFormat("\tE:{0}", w + 1);
}
// weights
sb.Append("\nWeights=");
if (_firstInputInitializationContent != null)
{
sb.AppendFormat("1");
}
if (NumTrees > 0)
{
if (_firstInputInitializationContent != null)
{
sb.Append("\t");
}
sb.AppendFormat("{0}", _trees[0].Weight);
}
for (int w = 1; w < NumTrees; ++w)
{
if (w > 0)
{
sb.Append("\t");
}
sb.Append(_trees[w].Weight);
}
sb.AppendFormat("\nBias={0}", Bias);
sb.Append("\nType=Linear");
// Add comments section with training parameters
int commentsWritten = AppendComments(sb, trainingParams);
if (appendFeatureGain)
{
var gainSummary = ToGainSummary(fmap, featureToID, NumTrees, includeZeroGainFeatures,
normalize: false, startingCommentNumber: commentsWritten);
sb.Append(gainSummary);
}
return(sb.ToString());
}