public override double CalculateScore(int variable, Varset parents, DoubleMap cache)
{
Varset parentsCp = new Varset(parents);
// TODO check if this violates the constraints
//if(constraints != NULL && !constraints.SatisfiesConstraints(variable, parents))
//{
//}
// check for prunning
double tVal = T(variable, parentsCp);
for (int x = 0; x < network.Size(); x++)
{
if (parentsCp.Get(x))
{
parentsCp.Set(x, false);
// TODO check the constraints
//if (invalidParents.Count > 0 && invalidParents.Contains(parentsCp.ToULong()))
//{
// parentsCp.Set(x);
// continue;
//}
double tmp = cache.ContainsKey(parentsCp.ToULong()) ? cache[parentsCp.ToULong()] : 0;
if (tmp + tVal > 0)
{
return 0;
}
parentsCp.Set(x, true);
}
}
double score = llc.Calculate(variable, parentsCp);
// structure penalty
score -= tVal * baseComplexityPenalty;
//Console.WriteLine("score: " + score);
return score;
}
public void Execute(string[] args)
{
// 引数のチェック
if (args.Length != 2)
{
// エラーメッセージ表示
Console.WriteLine("[エラー] 引数の数が不正です");
Environment.Exit(0);
}
// オプションのチェック
// 暫定的に初期値を代入
int rMin = 5; // The minimum number of records in the AD-tree nodes.
char delimiter = ',';
bool hasHeader = true;
string sf = "BDeu";
int maxParents = 0;
string constraintsFile = "";
int runningTime = -1;
int threadCount = 20;
bool prune = true;
string inputFile = args[0];
string outputFile = args[1];
double equivarentSampleSize = 1;
// csvファイルの読み込み
RecordFile recordFile = new RecordFile();
recordFile.ReadRecord(inputFile, hasHeader, delimiter);
//recordFile.Print();
// BayesianNetworkの初期化
BayesianNetwork network = new BayesianNetwork(recordFile);
// AD-Treeの初期化
ADTree adTree = new ADTree(rMin, network, recordFile);
// scoring functionの設定
sf = sf.ToLower();
if (maxParents > network.Size() || maxParents < 1)
{
maxParents = network.Size() - 1;
}
if (sf == "bic")
{
int maxParentCount = (int)Math.Log(2 * recordFile.Size() / Math.Log(recordFile.Size()));
if(maxParentCount < maxParents)
{
maxParents = maxParentCount;
}
}
else if (sf == "fnml") { }
else if (sf == "bdeu") { }
else
{
throw new ArgumentException("Invalid scoring function. Options are: 'BIC', 'fNML' or 'BDeu'.");
}
// TODO constraintsの実装
Constraints constraints = null;
if (constraintsFile.Length > 0)
{
constraints = Constraints.ParseConstraints(constraintsFile, network);
}
ScoringFunction scoringFunction = null;
List<double> ilogi = LogLikelihoodCalculator.GetLogCache(recordFile.Size());
LogLikelihoodCalculator llc = new LogLikelihoodCalculator(adTree, network, ilogi);
// TODO regretの実装
//std::vector<std::vector<float>*>* regret = scoring::getRegretCache(recordFile.size(), network.getMaxCardinality());
if (sf == "bic")
{
scoringFunction = new BICScoringFunction(network, recordFile, llc, constraints);
}
else if (sf == "fnml")
{
//
}
else if (sf == "bdeu")
{
scoringFunction = new BDeuScoringFunction(equivarentSampleSize, network, adTree, constraints);
}
ScoreCalculator scoreCalculator = new ScoreCalculator(scoringFunction, maxParents, network.Size(), runningTime, constraints);
// TODO Parallelクラスによる並列処理
Parallel.For(0, threadCount, i =>
{
for (int variable = 0; variable < network.Size(); variable++)
{
if(variable % threadCount != i)
{
continue;
}
Console.WriteLine("Thread: " + i + ", Variable: " + variable + ", Time: " + DateTime.Now);
DoubleMap sc = new DoubleMap();
scoreCalculator.CalculateScores(variable, sc);
int size = sc.Count;
Console.WriteLine("Thread: " + i + ", Variable: " + variable + ", Size before pruning: " + size + ", Time: " + DateTime.Now);
if (prune)
{
scoreCalculator.Prune(sc);
int prunedSize = sc.Count;
Console.WriteLine("Thread: " + i + ", Variable: " + variable + ", Size after pruning: " + prunedSize + ", Time: " + DateTime.Now);
}
string varFilename = outputFile + "." + variable;
StreamWriter writer = new StreamWriter(varFilename, false);
Variable var = network.Get(variable);
writer.Write("VAR " + var.Name + "\n");
writer.Write("META arity=" + var.GetCardinality() + "\n");
writer.Write("META values=");
for (int k = 0; k < var.GetCardinality(); k++)
{
writer.Write(var.GetValue(k) + " ");
}
writer.Write("\n");
foreach (KeyValuePair<ulong, double> kvp in sc)
{
Varset parentSet = new Varset(kvp.Key);
double s = kvp.Value;
writer.Write(s + " ");
for (int p = 0; p < network.Size(); p++)
{
if (parentSet.Get(p))
{
writer.Write(network.Get(p).Name + " ");
}
}
writer.Write("\n");
}
writer.Write("\n");
writer.Close();
}
});
// concatenate all of the files together
StreamWriter outFile = new StreamWriter(outputFile, false); // 修正が必要かも
// first, the header information
string header = "META pss_version = 0.1\nMETA input_file=" + inputFile + "\nMETA num_records=" + recordFile.Size() + "\n";
header += "META parent_limit=" + maxParents + "\nMETA score_type=" + sf + "\nMETA ess=" + equivarentSampleSize + "\n\n";
outFile.Write(header);
for (int variable = 0; variable < network.Size(); variable++)
{
string varFilename = outputFile + "." + variable;
StreamReader reader = new StreamReader(varFilename);
outFile.Write(reader.ReadToEnd());
reader.Close();
}
outFile.Close();
}
private void Calculate(ContingencyTableNode ct, ulong currentBase, ulong index, Dictionary<ulong, int> paCounts, int variable, Varset variables, int previousVariable, ref double score)
{
// if this is a leaf
if (ct.IsLeaf())
{
// update the instantiation count of this set of parents
int count = paCounts.ContainsKey(index) ? paCounts[index] : 0;
count += ct.Value;
paCounts[index] = count;
// update the score for this variable, parent instantiation
score += ilogi[ct.Value];
return;
}
// which actual variable are we looking at
int thisVariable = previousVariable + 1;
for (; thisVariable < network.Size(); thisVariable++)
{
if (variables.Get(thisVariable))
{
break;
}
}
// update the base and index if this is part of the parent set
ulong nextBase = currentBase;
if (thisVariable != variable)
{
nextBase *= (ulong)network.GetCardinality(thisVariable);
}
// recurse
for (int k = 0; k < network.GetCardinality(thisVariable); k++)
{
ContingencyTableNode child = ct.GetChild(k);
if (child != null)
{
ulong newIndex = index;
if (thisVariable != variable)
{
newIndex += currentBase * (ulong)k;
}
Calculate(child, nextBase, newIndex, paCounts, variable, variables, thisVariable, ref score);
}
}
}
private void CalculateScoresInternal(int variable, DoubleMap cache)
{
// calculate initial score
Varset empty = new Varset(variableCount + 1); // 注意: c++だと0
double score = scoringFunction.CalculateScore(variable, empty, cache);
if (score < 1)
{
cache[empty.ToULong()] = score;
}
int prunedCount = 0;
for (int layer = 1; layer <= maxParents && !outOfTime; layer++)
{
// DEBUG
Console.WriteLine("layer: " + layer + ", prunedCount: " + prunedCount);
Varset variables = new Varset(variableCount + 1); // 注意: c++だと0
for (int i = 0; i < layer; i++)
{
variables.Set(i, true);
}
Varset max = new Varset(variableCount);
max.Set(variableCount, true);
while (variables.LessThan(max) && !outOfTime)
{
if (!variables.Get(variable))
{
score = scoringFunction.CalculateScore(variable, variables, cache);
if (score < 0)
{
cache[variables.ToULong()] = score;
}
else
{
prunedCount++;
}
}
variables = variables.NextPermutation();
}
if (!outOfTime)
{
highestCompletedLayer = layer;
}
}
}
public override double CalculateScore(int variable, Varset parents, DoubleMap cache)
{
Scratch s = scratchSpace[variable];
// TODO check if this violates the constraints
//if (constraints != NULL && !constraints->satisfiesConstraints(variable, parents))
//{
// s->invalidParents.insert(parents);
// return 1;
//}
for (int x = 0; x < network.Size(); x++)
{
if (parents.Get(x))
{
parents.Set(x, false);
// TODO check the constraints
//if (invalidParents.Count > 0 && invalidParents.Contains(parents.ToULong()))
//{
// // we cannot say anything if we skipped this because of constraints
// parents.Set(x, true);
// continue;
//}
parents.Set(x, true);
}
}
Lg(parents, ref s);
Varset variables = new Varset(parents);
variables.Set(variable, true);
s.Score = 0;
ContingencyTableNode ct = adTree.MakeContab(variables);
Dictionary<ulong, int> paCounts = new Dictionary<ulong, int>();
Calculate(ct, 1, 0, paCounts, variables, -1, ref s);
// check constraints (Theorem 9 from de Campos and Ji '11)
// only bother if the alpha bound is okay
// check if can prune
if (s.Aij <= 0.8349)
{
double bound = -1.0 * ct.LeafCount * s.Lri;
// check each subset
for (int x = 0; x < network.Size(); x++)
{
if (parents.Get(x))
{
parents.Set(x, false);
// check the constraints
//if (s->invalidParents.find(parents) != s->invalidParents.end())
//{
// // we cannot say anything if we skipped this because of constraints
// VARSET_SET(parents, x);
// continue;
//}
double tmp = cache.ContainsKey(parents.ToULong()) ? cache[parents.ToULong()] : 0;
// if the score is larger (better) than the bound, then we can prune
if (tmp > bound)
{
return 0;
}
parents.Set(x, true);
}
}
}
foreach (KeyValuePair<ulong, int> kvp in paCounts)
{
s.Score += s.Lgij;
s.Score -= ScoreCalculator.GammaLn(s.Aij + kvp.Value);
}
//parents.Print();
//Console.WriteLine(variable);
return s.Score;
}
private void Lg(Varset parents, ref Scratch s)
{
int r = 1;
for (int pa = 0; pa < network.Size(); pa++)
{
if (parents.Get(pa))
{
r *= network.GetCardinality(pa);
}
}
s.Aij = ess / r;
s.Lgij = ScoreCalculator.GammaLn(s.Aij);
r *= network.GetCardinality(s.Variable);
s.Aijk = ess / r;
s.Lgijk = ScoreCalculator.GammaLn(s.Aijk);
}
private void Calculate(ContingencyTableNode ct, ulong currentBase, ulong index, Dictionary<ulong, int> paCounts, Varset variables, int previousVariable, ref Scratch s)
{
Varset variablesCp = new Varset(variables);
// if this is a leaf in the AD-tree
if (ct.IsLeaf())
{
// update the instantiation count of this set of parents
int count = paCounts.ContainsKey(index) ? paCounts[index] : 0;
count += ct.Value;
if (count > 0)
{
paCounts[index] = count;
// update the score for this variable, parent instantiation
double temp = ScoreCalculator.GammaLn(s.Aijk + ct.Value);
s.Score -= s.Lgijk;
s.Score += temp;
}
return;
}
// which actual variable are we looking at
int thisVariable = previousVariable + 1;
for (; thisVariable < network.Size(); thisVariable++)
{
if (variables.Get(thisVariable))
{
break;
}
}
// update the base and index if this is part of the parent set
ulong nextBase = currentBase;
if (thisVariable != s.Variable)
{
nextBase *= (ulong)network.GetCardinality(thisVariable);
}
// recurse
for (int k = 0; k < network.GetCardinality(thisVariable); k++)
{
ContingencyTableNode child = ct.GetChild(k);
if (child != null)
{
ulong newIndex = index;
if (thisVariable != s.Variable)
{
newIndex += currentBase * (ulong)k;
}
Calculate(child, nextBase, newIndex, paCounts, variables, thisVariable, ref s);
}
}
}
private double T(int variable, Varset parents)
{
double penalty = network.GetCardinality(variable) - 1;
for (int pa = 0; pa < network.Size(); pa++)
{
if (parents.Get(pa))
{
penalty *= network.GetCardinality(pa);
}
}
return penalty;
}
