public override Matrix InnerDecode(Sequence pSequence)
{
//Reset the network
netReset();
int numStates = pSequence.GetSize();
predicted_fnn = new int[numStates];
predicted_bnn = new int[numStates];
Matrix mForward = new Matrix(numStates, forwardRNN.L2);
Matrix mBackward = new Matrix(numStates, backwardRNN.L2);
Parallel.Invoke(() =>
{
//Computing forward RNN
for (int curState = 0; curState < numStates; curState++)
{
State state = pSequence.Get(curState);
forwardRNN.setInputLayer(state, curState, numStates, predicted_fnn);
forwardRNN.computeNet(state, mForward[curState]); //compute probability distribution
predicted_fnn[curState] = forwardRNN.GetBestOutputIndex();
forwardRNN.copyHiddenLayerToInput();
}
},
() =>
{
//Computing backward RNN
for (int curState = numStates - 1; curState >= 0; curState--)
{
State state = pSequence.Get(curState);
backwardRNN.setInputLayer(state, curState, numStates, predicted_bnn, false);
backwardRNN.computeNet(state, mBackward[curState]); //compute probability distribution
predicted_bnn[curState] = backwardRNN.GetBestOutputIndex();
backwardRNN.copyHiddenLayerToInput();
}
});
//Merge forward and backward
Matrix m = new Matrix(numStates, forwardRNN.L2);
for (int curState = 0; curState < numStates; curState++)
{
for (int i = 0; i < forwardRNN.L2; i++)
{
m[curState][i] = mForward[curState][i] + mBackward[curState][i];
}
}
return m;
}
public override int[] learnSentenceForRNNCRF(Sequence pSequence)
{
//Reset the network
int numStates = pSequence.GetSize();
int[] predicted = new int[numStates];
//Predict output
Matrix m = InnerDecode(pSequence);
ForwardBackward(numStates, m);
//Get the best result
predicted = new int[numStates];
for (int i = 0; i < numStates; i++)
{
State state = pSequence.Get(i);
logp += Math.Log10(m_Diff[i][state.GetLabel()]);
counter++;
predicted[i] = GetBestZIndex(i);
}
UpdateBigramTransition(pSequence);
netReset();
forwardRNN.m_Diff = m_Diff;
backwardRNN.m_Diff = m_Diff;
double[] output_fnn = new double[L2];
double[] output_bnn = new double[L2];
Parallel.Invoke(() =>
{
//Learn forward network
for (int curState = 0; curState < numStates; curState++)
{
// error propogation
State state = pSequence.Get(curState);
forwardRNN.setInputLayer(state, curState, numStates, predicted_fnn);
forwardRNN.computeNet(state, output_fnn); //compute probability distribution
forwardRNN.learnNet(state, curState);
forwardRNN.LearnBackTime(state, numStates, curState);
forwardRNN.copyHiddenLayerToInput();
}
},
() =>
{
for (int curState = numStates - 1; curState >= 0; curState--)
{
// error propogation
State state = pSequence.Get(curState);
backwardRNN.setInputLayer(state, curState, numStates, predicted_bnn, false);
backwardRNN.computeNet(state, output_bnn); //compute probability distribution
backwardRNN.learnNet(state, curState);
backwardRNN.LearnBackTime(state, numStates, curState);
backwardRNN.copyHiddenLayerToInput();
}
});
return predicted;
}
public override int[] PredictSentence(Sequence pSequence)
{
//Reset the network
int numStates = pSequence.GetSize();
int[] predicted = new int[numStates];
//Predict output
Matrix m = InnerDecode(pSequence);
//Merge forward and backward
for (int curState = 0; curState < numStates; curState++)
{
State state = pSequence.Get(curState);
//activation 2 --softmax on words
double sum = 0; //sum is used for normalization: it's better to have larger precision as many numbers are summed together here
for (int c = 0; c < forwardRNN.L2; c++)
{
if (m[curState][c] > 50) m[curState][c] = 50; //for numerical stability
if (m[curState][c] < -50) m[curState][c] = -50; //for numerical stability
double val = Math.Exp(m[curState][c]);
sum += val;
m[curState][c] = val;
}
for (int c = 0; c < forwardRNN.L2; c++)
{
m[curState][c] /= sum;
}
logp += Math.Log10(m[curState][state.GetLabel()]);
counter++;
predicted[curState] = GetBestOutputIndex(m, curState);
}
netReset();
double[] output = new double[L2];
//Learn forward network
for (int curState = 0; curState < numStates; curState++)
{
// error propogation
State state = pSequence.Get(curState);
forwardRNN.setInputLayer(state, curState, numStates, predicted_fnn);
forwardRNN.computeNet(state, output); //compute probability distribution
//Copy output result to forward net work's output
for (int i = 0; i < forwardRNN.L2; i++)
{
forwardRNN.neuOutput[i].ac = m[curState][i];
}
forwardRNN.learnNet(state, curState);
forwardRNN.LearnBackTime(state, numStates, curState);
forwardRNN.copyHiddenLayerToInput();
}
for (int curState = numStates - 1; curState >= 0; curState--)
{
// error propogation
State state = pSequence.Get(curState);
backwardRNN.setInputLayer(state, curState, numStates, predicted_bnn, false);
backwardRNN.computeNet(state, output); //compute probability distribution
//Copy output result to forward net work's output
for (int i = 0; i < backwardRNN.L2; i++)
{
backwardRNN.neuOutput[i].ac = m[curState][i];
}
backwardRNN.learnNet(state, curState);
backwardRNN.LearnBackTime(state, numStates, curState);
backwardRNN.copyHiddenLayerToInput();
}
return predicted;
}
public Sequence ExtractFeatures(Sentence sentence)
{
Sequence sequence = new Sequence();
int n = sentence.GetTokenSize();
List<string[]> features = sentence.GetFeatureSet();
//For each token, get its sparse and dense feature set according configuration and training corpus
sequence.SetSize(n);
for (int i = 0; i < n; i++)
{
State state = sequence.Get(i);
ExtractSparseFeature(i, n, features, state);
var spDenseFeature = ExtractDenseFeature(i, n, features);
state.SetDenseData(spDenseFeature);
}
return sequence;
}
