//compute grad of: sum{-log{P(y*|x,w)} + R(w)}
public double getGrad_BFGS(List <double> vecGrad, model m, dataSet X)
{
//-log(obj)
double error = 0;
int nbFeatures = _fGene.NCompleteFeature;
//int i = 0;
foreach (dataSeq im in X)
{
double err = 0;
err = getGrad(vecGrad, m, im, null);
error += err;
}
if (Global.reg != 0.0)
{
for (int f = 0; f < nbFeatures; f++)
{
vecGrad[f] += m.W[f] / (Global.reg * Global.reg);
}
}
if (Global.reg != 0.0)
{
List <double> tmpWeights = m.W;
double sum = listTool.squareSum(tmpWeights);
error += sum / (2.0 * Global.reg * Global.reg);
}
return(error);
}
public model(dataSet X, featureGenerator fGen)
{
_nTag = X.NTag;
_nHiddenState = Global.nHiddenStatePerTag * _nTag;
//default value is 0
if (Global.random == 0)
{
_w = new float[fGen.NCompleteFeature];
}
else if (Global.random == 1)
{
List <float> randList = randomDoubleTool.getRandomList_float(fGen.NCompleteFeature);
_w = randList.ToArray();
if (Global.tuneWeightInit)
{
if (Global.tmpW == null)
{
Global.tmpW = new float[_w.Length];
}
_w.CopyTo(Global.tmpW, 0);
}
}
else if (Global.random == 2)
{
_w = new float[fGen.NCompleteFeature];
Global.optimW.CopyTo(_w, 0);
}
else
{
throw new Exception("error");
}
}
public List <double> test_multi_mtl(List <dataSet> XXList, double iter, List <StreamWriter> swOutputList)
{
List <double> scoreList = new List <double>();
for (int i = 0; i < XXList.Count; i++)
{
dataSet X = XXList[i];
model m = _modelList[i];
List <double> scoreList_i;
if (Global.evalMetric == "tok.acc")
{
scoreList_i = decode_tokAcc(X, m, iter, swOutputList[i]);
}
else if (Global.evalMetric == "str.acc")
{
scoreList_i = decode_strAcc(X, m, iter, swOutputList[i]);
}
else if (Global.evalMetric == "f1")
{
scoreList_i = decode_fscore(X, m, iter, swOutputList[i]);
}
else
{
throw new Exception("error");
}
scoreList.Add(scoreList_i[0]);
}
return(scoreList);
}
//compute grad of: sum{-log{P(y*|x,w)}} + R(w)
public double getGrad_BFGS(List <double> g, model m, dataSet X)
{
double error = 0;
int nFeature = _fGene.NCompleteFeature;
foreach (dataSeq x in X)
{
double err = 0;
err = getGradCRF(g, m, x, null);
error += err;
}
if (Global.reg != 0.0)
{
for (int f = 0; f < nFeature; f++)
{
g[f] += m.W[f] / (Global.reg * Global.reg);
}
}
if (Global.reg != 0.0)
{
float[] tmpWeights = m.W;
double sum = arrayTool.squareSum(tmpWeights);
error += sum / (2.0 * Global.reg * Global.reg);
}
return(error);
}
public static void dataSplit(dataSet X, double v1, double v2, dataSet X1, dataSet X2)
{
if (v2 < v1)
{
throw new Exception("error");
}
X1.Clear();
X2.Clear();
X1.setDataInfo(X);
X2.setDataInfo(X);
int n1 = (int)(X.Count * v1);
int n2 = (int)(X.Count * v2);
for (int i = 0; i < X.Count; i++)
{
if (i >= n1 && i < n2)
{
X1.Add(X[i]);
}
else
{
X2.Add(X[i]);
}
}
}
static double test()
{
Console.WriteLine("reading test data...");
Global.swLog.WriteLine("reading test data...");
dataSet XX = new dataSet(Global.fFeatureTest, Global.fGoldTest);
Console.WriteLine("Done! test data size: {0}", XX.Count);
Global.swLog.WriteLine("Done! test data size: {0}", XX.Count);
//load model & feature files for testing
toolbox tb = new toolbox(XX, false);
Stopwatch timer = new Stopwatch();
timer.Start();
List <double> scoreList = tb.test(XX, 0);
timer.Stop();
double time = timer.ElapsedMilliseconds / 1000.0;
Global.timeList.Add(time);
double score = scoreList[0];
Global.scoreListList.Add(scoreList);
resSummarize.write();
return(score);
}
static void getSegments(dataSeq x, dataSet X2)
{
int rand = randomTool.getOneRandom_int(-100, 100);
if (rand <= 0)//forward
{
for (int node = 0; node < x.Count;)
{
int step = getStep();
dataSeq x2 = new dataSeq(x, node, step + Global.overlapLength);
X2.Add(x2);
node += step;
}
}
else//backward
{
for (int node = x.Count - 1; node >= 0;)
{
int step = getStep();
dataSeq x2 = new dataSeq(x, node, step + Global.overlapLength, false);
X2.Add(x2);
node -= step;
}
}
}
public static dataSet structSplit(dataSet X)
{
//make fractions
dataSet X2 = new dataSet(X.NTag, X.NFeature);
for (int t = 0; t < X.Count; t++)
{
dataSeq x = X[t];
if (Global.structReg && Global.miniSize != 0)
{
/*int step = getStep();
* //if (x.Count > 4 * step)
* if (x.Count > 4 * step && Global.segStep.ToString().Contains(".") == false)//divide x to 2 segments, then do fine segments
* {
* int rand = randomTool.getOneRandom_int(step, x.Count - step);
* dataSeq x1 = new dataSeq(x, 0, rand);
* dataSeq x2 = new dataSeq(x, rand, x.Count);
* getSegments(x1, X2);
* getSegments(x2, X2);
* }
* else*/
getSegments(x, X2);
}
else
{
X2.Add(x);
}
}
return(X2);
}
//this function can be called by train(), cv(), & richEdge.train()
public static double baseTrain(dataSet XTest, toolbox tb)
{
Global.reinitGlobal();
double score = 0;
for (int i = 0; i < Global.ttlIter; i++)
{
Global.glbIter++;
Stopwatch timer = new Stopwatch();
timer.Start();
double err = tb.train();
timer.Stop();
double time = timer.ElapsedMilliseconds / 1000.0;
Global.timeList.Add(time);
Global.errList.Add(err);
Global.diffList.Add(Global.diff);
List <double> scoreList = tb.test(XTest, i);
score = scoreList[0];
Global.scoreListList.Add(scoreList);
Global.swLog.WriteLine("iter{0} diff={1} train-time(sec)={2} {3}={4}%", Global.glbIter, Global.diff.ToString("e2"), time.ToString("f2"), Global.metric, score.ToString("f2"));
Global.swLog.WriteLine("------------------------------------------------");
Global.swLog.Flush();
Console.WriteLine("iter{0} diff={1} train-time(sec)={2} {3}={4}%", Global.glbIter, Global.diff.ToString("e2"), time.ToString("f2"), Global.metric, score.ToString("f2"));
//if (Global.diff < Global.convergeTol)
//break;
}
return(score);
}
public List <double> test(dataSet X, double iter)
{
string outfile = Global.outDir + Global.fOutput;
Global.swOutput = new StreamWriter(outfile);
List <double> scoreList;
if (Global.evalMetric == "tok.acc")
{
scoreList = decode_tokAcc(X, _model, iter);
}
else if (Global.evalMetric == "str.acc")
{
scoreList = decode_strAcc(X, _model, iter);
}
else if (Global.evalMetric == "f1")
{
scoreList = decode_fscore(X, _model, iter);
}
else
{
throw new Exception("error");
}
Global.swOutput.Close();
return(scoreList);
}
static double train(dataSet X = null, dataSet XX = null)
{
//load data
if (X == null && XX == null)
{
Console.WriteLine("\nreading training & test data...");
Global.swLog.WriteLine("\nreading training & test data...");
X = new dataSet(Global.fFeatureTrain, Global.fGoldTrain);
XX = new dataSet(Global.fFeatureTest, Global.fGoldTest);
dataSizeScale(X);
Console.WriteLine("data sizes (train, test): {0} {1}", X.Count, XX.Count);
Global.swLog.WriteLine("data sizes (train, test): {0} {1}", X.Count, XX.Count);
}
double score = 0;
toolbox tb = new toolbox(X, true);
score = baseTrain(XX, tb);
resSummarize.write();
//save model
if (Global.save == 1)
{
tb.Model.save(Global.fModel);
}
return(score);
}
//for multi-task
public toolbox(dataSet X, List <dataSet> XList, bool train = true)
{
if (train)//to train
{
_X = X;
_XList = XList;
_fGene = new featureGenerator(X);
_model = null;
_modelList = new List <model>();
for (int i = 0; i < Global.nTask; i++)
{
model m = new model(XList[i], _fGene);
_modelList.Add(m);
}
_inf = new inference(this);
_grad = new gradient(this);
initOptimizer();
}
else//to test
{
_X = X;
_XList = XList;
_model = null;
_modelList = new List <model>();
for (int i = 0; i < Global.nTask; i++)
{
model m = new model(Global.modelDir + i.ToString() + Global.fModel);
_modelList.Add(m);
}
_fGene = new featureGenerator(X);
_inf = new inference(this);
_grad = new gradient(this);
}
}
static void crossValidation()
{
//load data
Console.WriteLine("reading cross validation data...");
Global.swLog.WriteLine("reading cross validation data...");
List <dataSet> XList = new List <dataSet>();
List <dataSet> XXList = new List <dataSet>();
loadDataForCV(XList, XXList);
for (int i = 0; i < Global.nCV; i++)
{
Global.swLog.WriteLine("\n#validation={0}", i + 1);
Console.WriteLine("\n#validation={0}", i + 1);
if (Global.rawResWrite)
{
Global.swResRaw.WriteLine("% #validation={0}", i + 1);
}
dataSet Xi = XList[i];
toolbox tb = new toolbox(Xi);
baseTrain(XXList[i], tb);
resSummarize.write();
if (Global.rawResWrite)
{
Global.swResRaw.WriteLine();
}
}
}
public dataSet randomShuffle()
{
List <int> ri = randomTool <int> .getShuffledIndexList(this.Count);
dataSet X = new dataSet(this.NTag, this.NFeature);
foreach (int i in ri)
{
X.Add(this[i]);
}
return(X);
}
static void crossValidation()
{
//load data
Console.WriteLine("reading cross validation data...");
Global.swLog.WriteLine("reading cross validation data...");
List <dataSet> XList = new List <dataSet>();
List <dataSet> XXList = new List <dataSet>();
loadDataForCV(XList, XXList);
//start cross validation
foreach (double r in Global.regList)//do CV for each different regularizer r (sigma)
{
Global.swLog.WriteLine("\ncross validation. r={0}", r);
Console.WriteLine("\ncross validation. r={0}", r);
if (Global.rawResWrite)
{
Global.swResRaw.WriteLine("% cross validation. r={0}", r);
}
for (int i = 0; i < Global.nCV; i++)
{
Global.swLog.WriteLine("\n#validation={0}", i + 1);
Console.WriteLine("\n#validation={0}", i + 1);
if (Global.rawResWrite)
{
Global.swResRaw.WriteLine("% #validation={0}", i + 1);
}
Global.reg = r;
dataSet Xi = XList[i];
if (Global.runMode.Contains("rich"))
{
toolboxRich tb = new toolboxRich(Xi);
basicTrain(XXList[i], tb);
}
else
{
toolbox tb = new toolbox(Xi);
basicTrain(XXList[i], tb);
}
resSummarize.write();
if (Global.rawResWrite)
{
Global.swResRaw.WriteLine();
}
}
if (Global.rawResWrite)
{
Global.swResRaw.WriteLine();
}
}
}
//for train & test
public featureGenerator(dataSet X)
{
_nFeatureTemp = X.NFeatureTemp;
_nTag = X.NTag;
Global.swLog.WriteLine("feature templates: {0}", _nFeatureTemp);
int nNodeFeature = _nFeatureTemp * _nTag;
int nEdgeFeature = _nTag * _nTag;
_backoffEdge = nNodeFeature;
_nCompleteFeature = nNodeFeature + nEdgeFeature;
Global.swLog.WriteLine("complete features: {0}", _nCompleteFeature);
}
//for training
public featureGeneRich(dataSet X)
{
_nFeatureTemp = X.NFeatureTemp;
Global.swLog.WriteLine("feature templates: {0}", _nFeatureTemp);
_nTag = X.NTag;
_nState = X.NTag * Global.nHiddenStatePerTag;
int nNodeFeatures = _nFeatureTemp * _nState;
int nEdgeFeatures = _nFeatureTemp * _nState * _nState;
_backoffEdge = nNodeFeatures;
_nCompleteFeature = nNodeFeatures + nEdgeFeatures;
Global.swLog.WriteLine("complete features: {0}", _nCompleteFeature);
}
static void tuneWeightInit()
{
//tune good weight init for latent conditional models
if (Global.modelOptimizer.StartsWith("lsp"))
{
Console.WriteLine("\nreading training & test data...");
Global.swLog.WriteLine("\nreading training & test data...");
dataSet origX = new dataSet(Global.fFeatureTrain, Global.fGoldTrain);
dataSet X = new dataSet();
dataSet XX = new dataSet();
dataSplit(origX, Global.tuneSplit, X, XX);
//backup & change setting
int origTtlIter = Global.ttlIter;
Global.ttlIter = Global.iterTuneWeightInit;
Global.rawResWrite = false;
Global.tuneWeightInit = true;
Global.swTune.WriteLine("tuning weight initialization:");
Console.WriteLine("tuning weight initialization:");
double bestScore = 0;
int n = Global.nTuneRound;
for (int i = 0; i < n; i++)
{
Console.WriteLine("\ntuning-weight-init round {0} (a step before real training!):", i + 1);
train(X, XX);
double ttlScore = Global.ttlScore;
Global.swTune.WriteLine("score: {0}", ttlScore);
if (ttlScore > bestScore)
{
bestScore = ttlScore;
if (Global.optimW == null)
{
Global.optimW = new float[Global.tmpW.Length];
}
Global.tmpW.CopyTo(Global.optimW, 0);
}
else
{
Global.swTune.WriteLine("optimW no update.");
}
}
Global.swTune.Flush();
//recover setting
Global.random = 2;//2 means to init with optimal weights in toolbox
Global.ttlIter = origTtlIter;
Global.rawResWrite = true;
Global.tuneWeightInit = false;
}
}
//re-init global values and conduct training
static double reinitTrain(dataSet X = null, dataSet XX = null)
{
Global.reinitGlobal();
double score = 0;
if (Global.runMode.Contains("rich"))
{
score = richEdge.train(X, XX);
}
else
{
score = train(X, XX);
}
return(score);
}
public static void loadTestData_multi(List <dataSet> XXList)
{
XXList.Clear();
//load test data
Global.swLog.WriteLine("test data sizes (1, ..., T):");
for (int i = 0; i < Global.nTask; i++)
{
string dat_i = i.ToString() + Global.fFeatureTest;
string tag_i = i.ToString() + Global.fGoldTest;
dataSet Xtest = new dataSet(dat_i, tag_i);
Global.swLog.WriteLine(" " + Xtest.Count.ToString());
XXList.Add(Xtest);
}
Global.swLog.WriteLine();
}
public static double test()
{
dataSet X = new dataSet(Global.fFeatureTrain, Global.fGoldTrain);
dataSet XX = new dataSet(Global.fFeatureTest, Global.fGoldTest);
Global.swLog.WriteLine("data size (test): {0}", XX.Count);
//load model for testing
toolboxRich tb = new toolboxRich(X, false);
List <double> scoreList = tb.test(XX, 0);
double score = scoreList[0];
Global.scoreListList.Add(scoreList);
resSummarize.write();
return(score);
}
//string accuracy
public List <double> decode_strAcc(dataSet X, model m, double iter)
{
double xsize = X.Count;
double corr = 0;
//multi thread
List <dataSeqTest> X2 = new List <dataSeqTest>();
multiThreading(X, X2);
foreach (dataSeqTest x in X2)
{
//output tag results
if (Global.swOutput != null)
{
for (int i = 0; i < x._x.Count; i++)
{
Global.swOutput.Write(x._yOutput[i].ToString() + ",");
}
Global.swOutput.WriteLine();
}
List <int> goldTags = x._x.getTags();
bool ck = true;
for (int i = 0; i < x._x.Count; i++)
{
if (goldTags[i] != x._yOutput[i])
{
ck = false;
break;
}
}
if (ck)
{
corr++;
}
}
double acc = corr / xsize * 100.0;
Global.swLog.WriteLine("total-tag-strings={0} correct-tag-strings={1} string-accuracy={2}%", xsize, corr, acc);
List <double> scoreList = new List <double>();
scoreList.Add(acc);
return(scoreList);
}
static double train()
{
//load data
Console.WriteLine("\nreading training & test data...");
Global.swLog.WriteLine("\nreading training & test data...");
dataSet X, XX;
if (Global.runMode.Contains("tune"))//put "tune" related code here because train() could be sub-function of tune()
{
dataSet origX = new dataSet(Global.fFeatureTrain, Global.fGoldTrain);
X = new dataSet();
XX = new dataSet();
dataSplit(origX, Global.tuneSplit, X, XX);
}
else
{
X = new dataSet(Global.fFeatureTrain, Global.fGoldTrain);
XX = new dataSet(Global.fFeatureTest, Global.fGoldTest);
dataSizeScale(X);
}
Console.WriteLine("done! train/test data sizes: {0}/{1}", X.Count, XX.Count);
Global.swLog.WriteLine("done! train/test data sizes: {0}/{1}", X.Count, XX.Count);
double score = 0;
//start training
foreach (double r in Global.regList)//train on different r (sigma)
{
Global.reg = r;
Global.swLog.WriteLine("\nr: " + r.ToString());
Console.WriteLine("\nr: " + r.ToString());
if (Global.rawResWrite)
{
Global.swResRaw.WriteLine("\n%r: " + r.ToString());
}
toolbox tb = new toolbox(X, true);
score = basicTrain(XX, tb);
resSummarize.write();//summarize the results & output the summarized results
if (Global.save == 1)
{
tb.Model.save(Global.fModel);//save model as a .txt file
}
}
return(score);
}
//test3 mode in multi-task test: all models vote
public List <double> test3_multi_mtl(List <List <double> > vecList, List <dataSet> XXList, double iter, List <StreamWriter> swOutputList)
{
List <double> scoreList = new List <double>();
for (int i = 0; i < XXList.Count; i++)
{
dataSet X = XXList[i];
List <double> vec = MainClass.getVecFromX(X);
model m = new model(_modelList[0], false);
double ttlCos = 0;
for (int j = 0; j < vecList.Count; j++)
{
double cos = mathTool.cos(vecList[j], vec);
for (int k = 0; k < m.W.Count; k++)
{
m.W[k] += cos * _modelList[j].W[k];
ttlCos += cos;
}
}
for (int k = 0; k < m.W.Count; k++)
{
m.W[k] /= ttlCos;
}
List <double> scoreList_i;
if (Global.evalMetric == "tok.acc")
{
scoreList_i = decode_tokAcc(X, m, iter, swOutputList[i]);
}
else if (Global.evalMetric == "str.acc")
{
scoreList_i = decode_strAcc(X, m, iter, swOutputList[i]);
}
else if (Global.evalMetric == "f1")
{
scoreList_i = decode_fscore(X, m, iter, swOutputList[i]);
}
else
{
throw new Exception("error");
}
scoreList.Add(scoreList_i[0]);
}
return(scoreList);
}
public toolbox(dataSet X, bool train = true)
{
if (train)//for training
{
_X = X;
_fGene = new featureGenerator(X);
_model = new model(X, _fGene);
_inf = new inference(this);
initOptimizer();
}
else//for test
{
_X = X;
_model = new model(Global.fModel);
_fGene = new featureGenerator(X);
_inf = new inference(this);
}
}
public model(dataSet X, featureGenerator fGen)
{
_nTag = X.NTag;
//default value is 0
if (Global.random == 0)
{
_w = new float[fGen.NCompleteFeature];
}
else if (Global.random == 1)
{
List <float> randList = randomDoubleTool.getRandomList_float(fGen.NCompleteFeature);
_w = randList.ToArray();
}
else
{
throw new Exception("error");
}
}
//f-score
public List <double> decode_fscore(dataSet X, model m, double iter)
{
//multi thread
List <dataSeqTest> X2 = new List <dataSeqTest>();
multiThreading(X, X2);
List <string> goldTagList = new List <string>();
List <string> resTagList = new List <string>();
foreach (dataSeqTest x in X2)
{
string res = "";
foreach (int im in x._yOutput)
{
res += im.ToString() + ",";
}
resTagList.Add(res);
//output tag results
if (Global.swOutput != null)
{
for (int i = 0; i < x._yOutput.Count; i++)
{
Global.swOutput.Write(x._yOutput[i] + ",");
}
Global.swOutput.WriteLine();
}
List <int> goldTags = x._x.getTags();
string gold = "";
foreach (int im in goldTags)
{
gold += im.ToString() + ",";
}
goldTagList.Add(gold);
}
List <double> infoList = new List <double>();
List <double> scoreList = fscore.getFscore(goldTagList, resTagList, infoList);
Global.swLog.WriteLine("#gold-chunk={0} #output-chunk={1} #correct-output-chunk={2} precision={3} recall={4} f-score={5}", infoList[0], infoList[1], infoList[2], scoreList[1].ToString("f2"), scoreList[2].ToString("f2"), scoreList[0].ToString("f2"));
return(scoreList);
}
public static double train()
{
//load data
Console.WriteLine("\nreading training & test data...");
Global.swLog.WriteLine("\nreading training & test data...");
dataSet X, XX;
if (Global.runMode.Contains("tune"))
{
dataSet origX = new dataSet(Global.fFeatureTrain, Global.fGoldTrain);
X = new dataSet();
XX = new dataSet();
MainClass.dataSplit(origX, Global.tuneSplit, X, XX);
}
else
{
X = new dataSet(Global.fFeatureTrain, Global.fGoldTrain);
XX = new dataSet(Global.fFeatureTest, Global.fGoldTest);
MainClass.dataSizeScale(X);
}
Global.swLog.WriteLine("data sizes (train, test): {0} {1}", X.Count, XX.Count);
double score = 0;
foreach (double r in Global.regList)
{
Global.reg = r;
Global.swLog.WriteLine("\nr: " + r.ToString());
Console.WriteLine("\nr: " + r.ToString());
if (Global.rawResWrite)
{
Global.swResRaw.WriteLine("\n%r: " + r.ToString());
}
toolboxRich tb = new toolboxRich(X);
score = MainClass.basicTrain(XX, tb);
resSummarize.write();
//save model
if (Global.save == 1)
{
tb.Model.save(Global.fModel);
}
}
return(score);
}
public List <double> decode_fscore(dataSet XX, model m, double iter, StreamWriter swOutput)
{
int nTag = m.NTag;
double ttl = XX.Count;
List <string> goldTagList = new List <string>();
List <string> resTagList = new List <string>();
foreach (dataSeq x in XX)
{
//compute detected tags
List <int> tags = new List <int>();
double prob = _inf.decodeViterbi(m, x, tags);
string res = "";
foreach (int im in tags)
{
res += im.ToString() + ",";
}
resTagList.Add(res);
//output result tags
if (swOutput != null)
{
for (int i = 0; i < x.Count; i++)
{
swOutput.Write(tags[i] + ",");
}
swOutput.WriteLine();
}
List <int> goldTags = x.getTags();
string gold = "";
foreach (int im in goldTags)
{
gold += im.ToString() + ",";
}
goldTagList.Add(gold);
}
List <double> infoList = new List <double>();
List <double> scoreList = fscore.getFscore(goldTagList, resTagList, infoList);
return(scoreList);
}
public List <double> decode_strAcc(dataSet XX, model m, double iter, StreamWriter swOutput)
{
int nTag = m.NTag;
double ttl = XX.Count;
double correct = 0;
foreach (dataSeq x in XX)
{
//compute detected tags
List <int> tags = new List <int>();
double prob = _inf.decodeViterbi(m, x, tags);
//output result tags
if (swOutput != null)
{
for (int i = 0; i < x.Count; i++)
{
swOutput.Write(tags[i] + ",");
}
swOutput.WriteLine();
}
List <int> goldTags = x.getTags();
bool ck = true;
for (int i = 0; i < x.Count; i++)
{
if (goldTags[i] != tags[i])
{
ck = false;
break;
}
}
if (ck)
{
correct++;
}
}
double acc = correct / ttl;
List <double> scoreList = new List <double>();
scoreList.Add(acc);
return(scoreList);
}
