public void run(object param)
{
param pa = (param)param;
List <DataStep> x1 = pa.datastep;
List <Matrix> xpro = new List <Matrix>();
ForwdBackwdProp g = new ForwdBackwdProp(_train);
for (int i = 0; i < x1.Count; i++)
{
DataStep x = x1[i];
List <Matrix> add = new List <Matrix>();
for (int k = 0; k < 5; k++)
{
add.Add(Global.wordEmbedding[x.inputs[k]]);
}
List <Matrix> returnObj2 = Global.GRNNLayer1.activate(add, g);
List <Matrix> returnObj3 = Global.GRNNLayer2.activate(returnObj2, g);
List <Matrix> returnObj4 = Global.GRNNLayer3.activate(returnObj3, g);
List <Matrix> returnObj5 = Global.GRNNLayer4.activate(returnObj4, g);
xpro.Add(returnObj5[0]);
}
List <Matrix> returnObj6 = Global.upLSTMLayer.activate(xpro, g);
List <Matrix> returnObj7 = Global.upLSTMLayerr.activate(reverse(xpro), g);
List <Matrix> sum = new List <Matrix>();
for (int inde = 0; inde < returnObj6.Count(); inde++)
{
sum.Add(g.Add(returnObj6[inde], returnObj7[returnObj7.Count - inde - 1]));
}
for (int i = 0; i < returnObj6.Count; i++)
{
Matrix returnObj9 = Global.feedForwardLayer.Activate(sum[i], g);
double loss = LossSoftmax.getLoss(returnObj9, x1[i].goldOutput);
if (double.IsNaN(loss) || double.IsInfinity(loss))
{
Console.WriteLine("WARNING!!!");
Global.swLog.WriteLine("WARNING!!!");
pa.mre.Set();
return;
}
LossSoftmax.getGrad(returnObj9, x1[i].goldOutput);
}
g.backwardProp();
pa.mre.Set();
}
public static double runtestIteration(List <DataSeq> X, bool train)
{
double _total = 0, _prTotal = 0, _correct = 0;
double total = 0, accuracy = 0, recall = 0, f_score = 0;
double _total4 = 0, _prTotal4 = 0, _correct4 = 0, accuracy4 = 0, f_score4 = 0, recall4 = 0;
Global.threadList = new List <TrainThread>();
TrainThread runThread = new TrainThread(train);
List <ManualResetEvent> manualEvents = new List <ManualResetEvent>();
List <DataStep> temp = new List <DataStep>();
StreamWriter sw = new StreamWriter("data/temp/answer.txt");
string[] write_strings = new string[X.Count];
Parallel.ForEach(X, (ex) =>
{
param pa = new param(ex);
runThread.runtest(pa);
});
for (int i = 0; i < X.Count; i++)
{
sw.Write(X[i].write_string);
}
sw.Close();
accuracy = runThread._correct / runThread._prTotal;
recall = runThread._correct / runThread._total;
accuracy4 = runThread._correct4 / runThread._prTotal4;
recall4 = runThread._correct4 / runThread._total4;
f_score = 2 * accuracy * recall / (accuracy + recall);
f_score4 = 2 * accuracy4 * recall4 / (accuracy4 + recall4);
//Console.WriteLine("total windows: " + _total);
//Console.WriteLine("acc: " + (accuracy4 * 100).ToString("f3"));
//Console.WriteLine("recall: " + (recall4 * 100).ToString("f3"));
//Console.WriteLine("fscore: " + f_score4);
//Global.swLog.WriteLine("total windows: " + total);
//Global.swLog.WriteLine("acc: " + (accuracy4 * 100).ToString("f3"));
//Global.swLog.WriteLine("recall: " + (recall4 * 100).ToString("f3"));
//Global.swLog.WriteLine("fscore: " + f_score4);
//Global.swLog.Flush();
return(f_score4);
}
public static double runtrainIteration(List <DataSeq> X, List <DataSeq> Xtest, bool train, int iter)
{
List <DataSeq> x = new List <DataSeq>();
if (train)
{
x = shuffle(X);//shuffle every window (point)
}
TrainThread runThread = new TrainThread(train);
List <ManualResetEvent> manualEvents = new List <ManualResetEvent>();
List <DataStep> temp = new List <DataStep>();
int i = 0, j = 0;
int length = x.Count();
while (i < length)
{
if (i != 0 && i % 16 == 0)
{
testAccuracy1 = runtestIteration(Xtest, false);
if (testAccuracy <= testAccuracy1)
{
LSTMLayer.SerializeWordembedding("model//deepNetwork//embedding");
//LSTMLayer.SerializeBigramWordembedding();
Global.upLSTMLayer.saveLSTM("model//deepNetwork//lstmmodel.txt");
Global.upLSTMLayerr.saveLSTM("model//deepNetwork//lstmmodelr.txt");
Global.GRNNLayer1.saveGRNN("model//deepNetwork//grnnmodel1.txt");
Global.GRNNLayer2.saveGRNN("model//deepNetwork//grnnmodel1.txt");
Global.GRNNLayer3.saveGRNN("model//deepNetwork//grnnmodel1.txt");
Global.GRNNLayer4.saveGRNN("model//deepNetwork//grnnmodel1.txt");
Global.feedForwardLayer.saveFFmodel("model//deepNetwork//feedforwardmodel.txt");
testAccuracy = testAccuracy1;
}
Console.WriteLine("test f-score: {0}", (testAccuracy * 100).ToString("f3"));
Console.WriteLine("test1 f-score: {0}", (testAccuracy1 * 100).ToString("f3"));
Global.swLog.WriteLine("test f-score: {0}", (testAccuracy * 100).ToString("f3"));
Global.swLog.WriteLine("test f-score: {0}", (testAccuracy1 * 100).ToString("f3"));
}
for (int k = 0; k < Global.nThread && i < length; k++, i++)
{
ManualResetEvent mre = new ManualResetEvent(false);
param pa = new param(x[i]);
pa.mre = mre;
pa.datastep = x[i].datasteps;
manualEvents.Add(mre);
for (int m = 0; m < x[i].datasteps.Count; m++)
{
temp.Add(x[i].datasteps[m]);
}
ThreadPool.QueueUserWorkItem(new WaitCallback(runThread.run), pa);
}
WaitHandle.WaitAll(manualEvents.ToArray());
if (train)
{
UpdateWeight_rmProp(temp);
}
manualEvents.Clear();
temp.Clear();
}
return(runThread.accword / runThread.totalword);
}
public void runtest(object param)
{
param pa = (param)param;
List <DataStep> x1 = pa.datastep;
int[] ires4, igold4, wordindeis;
string str = "", str1 = "";
igold4 = new int[x1.Count];
ires4 = new int[x1.Count];
wordindeis = new int[x1.Count];
int index = 0, arraynum = 0;
ForwdBackwdProp g = new ForwdBackwdProp(_train);
int dim = 0;
dim = x1.Count;
Matrix[] xpro = new Matrix[dim];
List <int> ires_model = new List <int>();
//Parallel.For(0, temp.Count, i =>
for (int i = 0; i < x1.Count; i++)
{
List <Matrix> add = new List <Matrix>();
for (int k = 0; k < 5; k++)
{
add.Add(Global.wordEmbedding[x1[i].inputs[k]]);
}
List <Matrix> returnObj2 = Global.GRNNLayer1.activate(add, g);
List <Matrix> returnObj3 = Global.GRNNLayer2.activate(returnObj2, g);
List <Matrix> returnObj4 = Global.GRNNLayer3.activate(returnObj3, g);
List <Matrix> returnObj5 = Global.GRNNLayer4.activate(returnObj4, g);
xpro[i] = returnObj5[0];
}//);
List <Matrix> returnObj6 = Global.upLSTMLayer.activate(xpro.ToList(), g);
List <Matrix> returnObj7 = Global.upLSTMLayerr.activate(reverse(xpro.ToList()), g);
List <Matrix> sum = new List <Matrix>();
for (int inde = 0; inde < returnObj6.Count(); inde++)
{
sum.Add(g.Add(returnObj6[inde], returnObj7[returnObj7.Count - inde - 1]));
}
for (int i = 0; i < xpro.Length; i++)
{
Matrix returnObj9 = Global.feedForwardLayer.Activate(sum[i], g);
igold4[i] = LossSoftmax.getMax(x1[i].goldOutput);
ires4[i] = LossSoftmax.getMax(returnObj9);
}
//);
//fscore.backprocess(wordindeis, ires4);
pa.seq.write_string = "BOS O O" + "\n";
//pa.sw.WriteLine("BOS O O");
for (int i = 0; i < ires4.Count(); i++)
{
pa.seq.write_string += (Global.word[x1[i].wordindex] + " " + ires4[i] + " " + igold4[i] + "\n");
}
pa.seq.write_string += ("EOS O O" + "\n");
pa.seq.write_string += "\n";
List <string> res = fscore.getChunks4(ires4);
str1 = fscore.calcorrect(fscore.getChunks4(igold4), res);
lock (thislock)
{
string[] strs1 = str1.Split();
_total4 += Int32.Parse(strs1[0]);
_prTotal4 += Int32.Parse(strs1[1]);
_correct4 += Int32.Parse(strs1[2]);
}
}
