public Sequence BuildSequence(Sentence sentence)
{
var n = sentence.TokensList.Count;
var sequence = new Sequence(n);
//For each token, get its sparse and dense feature set according configuration and training corpus
for (var i = 0; i < n; i++)
{
var state = sequence.States[i];
ExtractSparseFeature(i, n, sentence.TokensList, state);
}
if (preTrainType == RNNSharp.PRETRAIN_TYPE.AutoEncoder)
{
var outputs = autoEncoder.ComputeTopHiddenLayerOutput(sentence);
for (var i = 0; i < n; i++)
{
var state = sequence.States[i];
state.DenseFeature = new SingleVector(outputs[i]);
}
}
else
{
for (var i = 0; i < n; i++)
{
var state = sequence.States[i];
state.DenseFeature = ExtractDenseFeature(i, n, sentence.TokensList);
}
}
return(sequence);
}
private void ExtractSourceSentenceFeature(RNNDecoder decoder, Sequence srcSequence, int targetSparseFeatureSize, out double[] srcHiddenAvgOutput, out Dictionary <int, float> srcSparseFeatures)
{
List <double[]> srcOutputs = decoder.ComputeTopHiddenLayerOutput(srcSequence);
srcHiddenAvgOutput = new double[srcOutputs[0].Length];
for (int i = 0; i < srcOutputs[0].Length; i++)
{
srcHiddenAvgOutput[i] = (srcOutputs[0][i] + srcOutputs[srcOutputs.Count - 1][i]) / 2.0;
}
srcSparseFeatures = new Dictionary <int, float>();
for (int i = 0; i < srcSequence.States.Length; i++)
{
foreach (KeyValuePair <int, float> kv in srcSequence.States[i].SparseFeature)
{
int srcSparseFeatureIndex = kv.Key + targetSparseFeatureSize;
if (srcSparseFeatures.ContainsKey(srcSparseFeatureIndex) == false)
{
srcSparseFeatures.Add(srcSparseFeatureIndex, kv.Value);
}
else
{
srcSparseFeatures[srcSparseFeatureIndex] += kv.Value;
}
}
}
}
/// <summary>
/// Extract features from source sequence
/// </summary>
/// <param name="decoder"></param>
/// <param name="srcSequence"></param>
/// <param name="targetSparseFeatureSize"></param>
/// <param name="srcHiddenAvgOutput"></param>
/// <param name="srcSparseFeatures"></param>
private void ExtractSourceSentenceFeature(RNNDecoder decoder, Sequence srcSequence, int targetSparseFeatureSize,
out float[] srcHiddenAvgOutput, out Dictionary <int, float> srcSparseFeatures)
{
//Extract dense features from source sequence
var srcOutputs = decoder.ComputeTopHiddenLayerOutput(srcSequence);
int srcSequenceDenseFeatureSize = srcOutputs[0].Length;
int srcSequenceLength = srcOutputs.Count - 1;
srcHiddenAvgOutput = new float[srcSequenceDenseFeatureSize * 2];
var j = 0;
float[] srcOutputForward = srcOutputs[0];
float[] srcOutputBackward = srcOutputs[srcSequenceLength];
while (j < srcSequenceDenseFeatureSize - Vector <float> .Count)
{
var vForward = new Vector <float>(srcOutputForward, j);
var vBackward = new Vector <float>(srcOutputBackward, j);
vForward.CopyTo(srcHiddenAvgOutput, j);
vBackward.CopyTo(srcHiddenAvgOutput, srcSequenceDenseFeatureSize + j);
j += Vector <float> .Count;
}
while (j < srcSequenceDenseFeatureSize)
{
srcHiddenAvgOutput[j] = srcOutputForward[j];
srcHiddenAvgOutput[srcSequenceDenseFeatureSize + j] = srcOutputBackward[j];
j++;
}
//Extract sparse features from source sequence
srcSparseFeatures = new Dictionary <int, float>();
for (var i = 0; i < srcSequence.States.Length; i++)
{
foreach (var kv in srcSequence.States[i].SparseFeature)
{
var srcSparseFeatureIndex = kv.Key + targetSparseFeatureSize;
if (srcSparseFeatures.ContainsKey(srcSparseFeatureIndex) == false)
{
srcSparseFeatures.Add(srcSparseFeatureIndex, kv.Value);
}
else
{
srcSparseFeatures[srcSparseFeatureIndex] += kv.Value;
}
}
}
}
