//multi-task learning
public static void train_multi_mtl(List <dataSet> XXList, toolbox tb)
{
Global.reinitGlobal();
for (double iter = 0; iter < Global.ttlIter; iter++)
{
Global.glbIter++;
Stopwatch timer = new Stopwatch();
timer.Start();
double error = tb.train_multi();
timer.Stop();
double time = timer.ElapsedMilliseconds / 1000.0;
Global.swLog.WriteLine("Training used time (second): " + time.ToString());
//evaluate
List <double> scoreList = tb.test_multi_mtl(XXList, iter, Global.swOutputList);
for (int i = 0; i < Global.nTask; i++)
{
Global.scoreTaskList_multi[i].Add(scoreList[i]);
}
Global.timeList_multi[Global.glbIter - 1] += time;
Global.errorList_multi[Global.glbIter - 1] += error;
if (iter > 30 && Global.diff > 0 && Global.diff < Global.convergeTol)
{
break;
}
}
//save model
if (Global.save == 1)
{
for (int i = 0; i < Global.nTask; i++)
{
tb.ModelList[i].save(Global.modelDir + i.ToString() + Global.fModel);
}
}
}
static void multiTask()
{
//train
if (Global.runMode.Contains("train"))
{
//load data
List <dataSet> XList = new List <dataSet>();
List <dataSet> XXList = new List <dataSet>();
dataSet X = new dataSet();
loadData_multi(XList, X, XXList);
toolbox toolbox;
//single-task training in multi-task framework: each task has its own independent train & test data
if (Global.mt_singleTrain)
{
foreach (double r in Global.regList)//experiments for each different regularizer value
{
Global.swResRaw.WriteLine("\n%single-task! r: {0}", r);
Console.WriteLine("\nsingle-task! r: {0}", r);
for (int i = 0; i < Global.nTask; i++)
{
Global.swLog.WriteLine("\nsingle-task! #task, r: " + (i + 1).ToString() + "," + r.ToString());
Console.WriteLine("\nsingle-task! #task, r: " + (i + 1).ToString() + "," + r.ToString());
Global.reg = r;
dataSet Xi = XList[i];
toolbox = new toolbox(Xi);
train_multi_single(XXList, toolbox, i);
}
resProcess.write_multi();
}
Global.swResRaw.WriteLine();
}
//merged training in multi-task framework: merge all training data to train a unified model
if (Global.mt_mergeTrain)
{
foreach (double r in Global.regList)//experiments for each different regularizer value
{
Global.reg = r;
Global.swLog.WriteLine("\nmerged-task! r: " + r.ToString());
Console.WriteLine("\nmerged-task! r: " + r.ToString());
Global.swResRaw.WriteLine("\n%merged-task! r: " + r.ToString());
toolbox = new toolbox(X);
train_multi_merge(XXList, toolbox);
resProcess.write_multi();
}
Global.swResRaw.WriteLine();
}
//multi-task learning
if (Global.mt_mtTrain)
{
foreach (double r in Global.regList)//experiments for each different regularizer value
{
Global.reg = r;
foreach (double cFactor in Global.cFactors)//experiments for each different C value (see Eq. 18 & 19 of [Sun+ TKDE 2013] for the definition of C)
{
Global.C = cFactor;
Global.swLog.WriteLine("\n%multi-task! reg, rate0, C, kernel: {0},{1},{2},{3}", Global.reg, Global.rate0, Global.C, Global.simiMode);
Global.swSimi.WriteLine("\n%multi-task! reg, rate0, C, kernel: {0},{1},{2},{3}", Global.reg, Global.rate0, Global.C, Global.simiMode);
Global.swResRaw.WriteLine("\n%multi-task! reg, rate0, C, kernel: {0},{1},{2},{3}", Global.reg, Global.rate0, Global.C, Global.simiMode);
Console.WriteLine("\nmulti-task! reg, rate0, C, kernel: {0},{1},{2},{3}", Global.reg, Global.rate0, Global.C, Global.simiMode);
toolbox = new toolbox(X, XList);
train_multi_mtl(XXList, toolbox);
resProcess.write_multi();
}
}
Global.swResRaw.WriteLine();
}
}
else if (Global.runMode.Contains("test1"))//normal test
{
//load data
List <dataSet> XList = new List <dataSet>();
List <dataSet> XXList = new List <dataSet>();
dataSet X = new dataSet();
loadData_multi(XList, X, XXList);
//load model etc.
toolbox tb = new toolbox(X, XList, false);
if (Global.mt_mergeTrain)//multi_merge
{
List <double> scoreList = tb.test_multi_merge(XXList, 0, Global.swOutputList);
for (int i = 0; i < Global.nTask; i++)
{
Global.scoreTaskList_multi[i].Add(scoreList[i]);
}
resProcess.write_multi();
}
else//multi_single or multi_mtl: they have the same testing schema
{
List <double> scoreList = tb.test_multi_mtl(XXList, 0, Global.swOutputList);
for (int i = 0; i < Global.nTask; i++)
{
Global.scoreTaskList_multi[i].Add(scoreList[i]);
}
resProcess.write_multi();
}
}
else if (Global.runMode.Contains("test2"))//for multi_mtl: test a new task via choosing the most similar model
{
//load data
List <dataSet> XList = new List <dataSet>();
List <dataSet> XXList = new List <dataSet>();
dataSet X = new dataSet();
loadData_multi(XList, X, XXList);
//get vectors
List <List <double> > vecList = new List <List <double> >();
foreach (dataSet Xi in XList)
{
List <double> vec = getVecFromX(Xi);
vecList.Add(vec);
}
//load model & test
toolbox tb = new toolbox(X, XList, false);
List <double> scoreList = tb.test2_multi_mtl(vecList, XXList, 0, Global.swOutputList);
for (int i = 0; i < Global.nTask; i++)
{
Global.scoreTaskList_multi[i].Add(scoreList[i]);
}
resProcess.write_multi();
}
else if (Global.runMode.Contains("test3"))//for multi_mtl: test a new task via voted-test based on all models, i.e., the OMT-SBD method described in Section 4.4 of [Sun+ TKDE 2013]
{
//load data
List <dataSet> XList = new List <dataSet>();
List <dataSet> XXList = new List <dataSet>();
dataSet X = new dataSet();
loadData_multi(XList, X, XXList);
//get vectors
List <List <double> > vecList = new List <List <double> >();
foreach (dataSet Xi in XList)
{
List <double> vec = getVecFromX(Xi);
vecList.Add(vec);
}
//load model & test
toolbox tb = new toolbox(X, XList, false);
List <double> scoreList = tb.test3_multi_mtl(vecList, XXList, 0, Global.swOutputList);
for (int i = 0; i < Global.nTask; i++)
{
Global.scoreTaskList_multi[i].Add(scoreList[i]);
}
resProcess.write_multi();
}
else
{
throw new Exception("error");
}
}