public int[][] DecodeNBestCRF(Sequence seq, int N)
{
//ys contains the output of RNN for each word
Matrix <double> ys = PredictSentence(seq, RunningMode.Test);
int n = seq.States.Length;
int K = L2;
Matrix <double> STP = CRFTagTransWeights;
PAIR <int, int>[, ,] vPath = new PAIR <int, int> [n, K, N];
int DUMP_LABEL = -1;
double[,] vPreAlpha = new double[K, N];
double[,] vAlpha = new double[K, N];
int nStartTagIndex = 0;
int nEndTagIndex = 0;
double MIN_VALUE = double.MinValue;
//viterbi algorithm
for (int i = 0; i < K; i++)
{
for (int j = 0; j < N; j++)
{
vPreAlpha[i, j] = MIN_VALUE;
vPath[0, i, j] = new PAIR <int, int>(DUMP_LABEL, 0);
}
}
vPreAlpha[nStartTagIndex, 0] = ys[0][nStartTagIndex];
vPath[0, nStartTagIndex, 0].first = nStartTagIndex;
AdvUtils.PriorityQueue <double, PAIR <int, int> > q = new AdvUtils.PriorityQueue <double, PAIR <int, int> >();
for (int t = 1; t < n; t++)
{
for (int j = 0; j < K; j++)
{
while (q.Count() > 0)
{
q.Dequeue();
}
double _stp = STP[j][0];
double _y = ys[t][j];
for (int k = 0; k < N; k++)
{
double score = vPreAlpha[0, k] + _stp + _y;
q.Enqueue(score, new PAIR <int, int>(0, k));
}
for (int i = 1; i < K; i++)
{
_stp = STP[j][i];
for (int k = 0; k < N; k++)
{
double score = vPreAlpha[i, k] + _stp + _y;
if (score <= q.Peek().Key)
{
break;
}
q.Dequeue();
q.Enqueue(score, new PAIR <int, int>(i, k));
}
}
int idx = N - 1;
while (q.Count() > 0)
{
vAlpha[j, idx] = q.Peek().Key;
vPath[t, j, idx] = q.Peek().Value;
idx--;
q.Dequeue();
}
}
vPreAlpha = vAlpha;
vAlpha = new double[K, N];
}
//backtrace to get the n-best result path
int[][] vTagOutput = new int[N][];
for (int i = 0; i < N; i++)
{
vTagOutput[i] = new int[n];
}
for (int k = 0; k < N; k++)
{
vTagOutput[k][n - 1] = nEndTagIndex;
PAIR <int, int> decision = new PAIR <int, int>(nEndTagIndex, k);
for (int t = n - 2; t >= 0; t--)
{
vTagOutput[k][t] = vPath[t + 1, decision.first, decision.second].first;
decision = vPath[t + 1, decision.first, decision.second];
}
}
return(vTagOutput);
}
public int[][] DecodeNBestCRF(Sequence seq, int N)
{
//ys contains the output of RNN for each word
Matrix<double> ys = ProcessSequence(seq, RunningMode.Test);
int n = seq.States.Length;
int K = OutputLayer.LayerSize;
Matrix<double> STP = CRFTagTransWeights;
PAIR<int, int>[,,] vPath = new PAIR<int, int>[n, K, N];
int DUMP_LABEL = -1;
double[,] vPreAlpha = new double[K, N];
double[,] vAlpha = new double[K, N];
int nStartTagIndex = 0;
int nEndTagIndex = 0;
double MIN_VALUE = double.MinValue;
//viterbi algorithm
for (int i = 0; i < K; i++)
{
for (int j = 0; j < N; j++)
{
vPreAlpha[i, j] = MIN_VALUE;
vPath[0, i, j] = new PAIR<int, int>(DUMP_LABEL, 0);
}
}
vPreAlpha[nStartTagIndex, 0] = ys[0][nStartTagIndex];
vPath[0, nStartTagIndex, 0].first = nStartTagIndex;
AdvUtils.PriorityQueue<double, PAIR<int, int>> q = new AdvUtils.PriorityQueue<double, PAIR<int, int>>();
for (int t = 1; t < n; t++)
{
for (int j = 0; j < K; j++)
{
while (q.Count() > 0)
q.Dequeue();
double _stp = STP[j][0];
double _y = ys[t][j];
for (int k = 0; k < N; k++)
{
double score = vPreAlpha[0, k] + _stp + _y;
q.Enqueue(score, new PAIR<int, int>(0, k));
}
for (int i = 1; i < K; i++)
{
_stp = STP[j][i];
for (int k = 0; k < N; k++)
{
double score = vPreAlpha[i, k] + _stp + _y;
if (score <= q.Peek().Key)
break;
q.Dequeue();
q.Enqueue(score, new PAIR<int, int>(i, k));
}
}
int idx = N - 1;
while (q.Count() > 0)
{
vAlpha[j, idx] = q.Peek().Key;
vPath[t, j, idx] = q.Peek().Value;
idx--;
q.Dequeue();
}
}
vPreAlpha = vAlpha;
vAlpha = new double[K, N];
}
//backtrace to get the n-best result path
int[][] vTagOutput = new int[N][];
for (int i = 0; i < N; i++)
{
vTagOutput[i] = new int[n];
}
for (int k = 0; k < N; k++)
{
vTagOutput[k][n - 1] = nEndTagIndex;
PAIR<int, int> decision = new PAIR<int, int>(nEndTagIndex, k);
for (int t = n - 2; t >= 0; t--)
{
vTagOutput[k][t] = vPath[t + 1, decision.first, decision.second].first;
decision = vPath[t + 1, decision.first, decision.second];
}
}
return vTagOutput;
}
