public ViterbiForGlobalLinearModel(WeightVector weightVector, Tags tags)
{
WeightVector = weightVector;
Tags = tags;
Pi = new List<Dictionary<string, double>>();
Bp = new List<Dictionary<string, string>>();
}
public ForwardBackwordAlgo(List<string> inputSentence, WeightVector wc, List<string> tagList)
{
_inputSentence = inputSentence;
_wc = wc;
_tagList = tagList;
_alphaDictionary = new Dictionary<int, Dictionary<string, double>>();
_betaDictionary = new Dictionary<int, Dictionary<string, double>>();
_uDictionary = new Dictionary<int, Dictionary<string, double>>();
UabDictionary = new Dictionary<int, Dictionary<string, double>>();
Z = 0;
_weightedFeaturesum = new WeightedFeatureSum(wc, inputSentence, true);
cList = new List<double>(_inputSentence.Count);
dList = new List<double>(_inputSentence.Count);
_useScaling = true;
_useLog = false;
_twoGramsList = new string[4];
var ngramTags = new Tags(_tagList);
int index = 0;
foreach (var ngram in ngramTags.GetNGramTags(2))
{
if (index >= _twoGramsList.Length)
{
Array.Resize(ref _twoGramsList, index + 1);
}
string[] split = ngram.Split(new[] { ':' });
_twoGramsList[index] = split[0] + "@#" + split[1];
index++;
}
}
public ComputeGradient(List<List<string>> inputSentence, List<List<string>> tagsList,
List<string> tagList, double lambda, double learningParam, FeatureCache cache, WriteModel logger)
{
Logger = logger;
_inputSentence = inputSentence;
_outputTagsList = tagsList;
_tagList = tagList;
_lambda = lambda;
_learningParam = learningParam;
_cache = cache;
forwardBackwordAlgos = new List<ForwardBackwordAlgo>();
_weightVector = null;
_twoGramsList = new string[4];
_twoGramPair = new KeyValuePair<string, string>[4];
var ngramTags = new Tags(_tagList);
int index = 0;
foreach (var ngram in ngramTags.GetNGramTags(2))
{
if (index >= _twoGramsList.Length)
{
Array.Resize(ref _twoGramsList, index+1);
Array.Resize(ref _twoGramPair, index + 1);
}
string[] split = ngram.Split(new[] { ':' });
_twoGramsList[index] = split[0] +"@#"+ split[1];
_twoGramPair[index] = new KeyValuePair<string, string>(split[0], split[1]);
index++;
}
}
public void Setup(bool debug)
{
var readModel = new ReadModel(InputModelFile);
var temp = new ReadModel(string.Concat(InputModelFile, ".featuresToK"));
_weightVector = new WeightVector(temp.GetFeatureToKdDictionary());
foreach (var pair in readModel.ModelIterator())
{
_weightVector.Add(pair);
}
_tags = new Tags(_tagList);
_viterbiForGlobalLinearModel = new ViterbiForGlobalLinearModel(_weightVector, _tags);
// read input file in a class and per line iterator.
var inputData = new ReadInputData(InputTestFile);
var writeModel = new WriteModel(_outputTestFile);
foreach (var line in inputData.GetSentence())
{
List<string> debugList;
var outputTags = _viterbiForGlobalLinearModel.Decode(line, debug, out debugList);
if (debug)
{
writeModel.WriteDataWithTagDebug(line, outputTags, debugList);
}
else
{
writeModel.WriteDataWithTag(line, outputTags);
}
}
writeModel.Flush();
}
public ThreadInfoObject(ComputeGradient cg, int start, int end, WeightVector wc, ManualResetEvent resetEvent)
{
Gradient = cg;
Start = start;
End = end;
NewWeightVector = wc;
ResetEvent = resetEvent;
}
public ComputeGradient(List<List<string>> inputSentence, List<List<string>> tagsList,
List<string> tagList, double lambda, FeatureCache cache)
{
_inputSentence = inputSentence;
_outputTagsList = tagsList;
_tagList = tagList;
_lambda = lambda;
_cache = cache;
forwardBackwordAlgos = new List<ForwardBackwordAlgo>();
_weightVector = null;
}
public void ReMappingFromWeightVector(WeightVector weightVector, bool normalize = true)
{
var newDictKtoF = new Dictionary<int, string>();
var newDictFtoK = new Dictionary<string, int>();
//var weightDict = new Dictionary<int, double>();
//int k = 0;
//var elements = from element in weightVector.WeightArray
// orderby element descending
// where element > 1
// select element;
//var sortedDictionary = from pair in weightVector.WDictionary
// where Math.Abs(pair.Value) > 1
// orderby Math.Abs(pair.Value) descending
// select pair;
const double limit = 0;
var newWeights = new double[weightVector.FeatureCount];
int featureCount = 0;
Array.Clear(newWeights, 0, newWeights.Length);
double max = 0;
for (int i = 0; i < weightVector.FeatureCount; i++)
{
if (max < Math.Abs(weightVector.WeightArray[i]))
{
max = Math.Abs(weightVector.WeightArray[i]);
}
if (Math.Abs(weightVector.WeightArray[i]) > limit)
{
newWeights[featureCount] = weightVector.WeightArray[i];
var feature = DictKToFeatures[i];
newDictFtoK[feature] = featureCount;
newDictKtoF[featureCount] = feature;
featureCount++;
}
}
//if (normalize)
//{
// for (int i = 0; i < featureCount; i++)
// {
// newWeights[i] /= max;
// }
//}
weightVector.WeightArray = newWeights;
weightVector.FeatureCount = featureCount;
DictFeaturesToK = weightVector.FeatureKDictionary = newDictFtoK;
DictKToFeatures = newDictKtoF;
if (normalize)
{
weightVector.AvgNormalize();
}
}
public Perceptron(string inputFile, string outputFile, List<string> tagList)
{
_inputFile = inputFile;
_outputFile = outputFile;
var tags = new Tags(tagList);
MapFeatures = new MapFeaturesToK(inputFile, string.Concat(outputFile, ".featuresToK"), tagList);
MapFeatures.StartMapping();
WeightVector = new WeightVector(MapFeatures.DictFeaturesToK, MapFeatures.FeatureCount);
_viterbiForGlobalLinearModel = new ViterbiForGlobalLinearModel(WeightVector, tags);
InputSentences = new List<List<string>>();
TagsList = new List<List<string>>();
ReadInputs();
}
public ForwardBackwordAlgo(List<string> inputSentence, WeightVector wc, List<string> tagList)
{
_inputSentence = inputSentence;
_wc = wc;
_tagList = tagList;
_tags = new Tags(tagList);
_alphaDictionary = new Dictionary<int, Dictionary<string, double>>();
_betaDictionary = new Dictionary<int, Dictionary<string, double>>();
_uDictionary = new Dictionary<int, Dictionary<string, double>>();
UabDictionary = new Dictionary<int, Dictionary<string, double>>();
Z = 0;
_weightedFeaturesum = new WeightedFeatureSum(wc, inputSentence, true);
}
public Perceptron(List<string> inputFiles, string outputFile, List<string> tagList, bool useAvg = false)
{
_outputFile = outputFile;
_useAvg = useAvg;
var tags = new Tags(tagList);
MapFeatures = new MapFeaturesToK(string.Concat(outputFile, ".featuresToK"), tagList);
MapFeatures.StartMapping(inputFiles);
WeightVector = new WeightVector(MapFeatures.DictFeaturesToK, MapFeatures.FeatureCount);
AvgWeightVector = new WeightVector(MapFeatures.DictFeaturesToK, MapFeatures.FeatureCount);
_viterbiForGlobalLinearModel = new ViterbiForGlobalLinearModel(WeightVector, tags);
InputSentences = new List<List<string>>();
TagsList = new List<List<string>>();
//ReadInputs();
}
public void Init()
{
var readModel = new ReadModel(InputModelFile + ".preceptron");
var temp = new ReadModel(string.Concat(InputModelFile, ".featuresToK"));
var dict = temp.GetFeatureToKdDictionary();
_weightVector = new WeightVector(dict, dict.Count);
foreach (var pair in readModel.ModelIterator())
{
_weightVector.Add(pair);
}
_tags = new Tags(_tagList);
ViterbiForGLM = new ViterbiForGlobalLinearModel(_weightVector, _tags);
}
public void ComputeRange(int start, int end, WeightVector newWeightVector, int threadIndex = 0)
{
for (var k = start; k < end; k++)
{
if (k % 100 == 0)
{
Console.WriteLine(DateTime.Now + "threadIndex: " + threadIndex +
" running iteration for k " + k);
}
var wk = Compute(k);
if (double.IsNaN(wk) || double.IsInfinity(wk))
{
Logger.WriteLine("k: "+ k + "wk is infiity of nana"+ wk);
Logger.Flush(false);
}
newWeightVector.SetKey(k, wk);
}
}
public WeightVector RunIterations(WeightVector weightVector, int iterationCount)
{
for (int iter = 0; iter < iterationCount; iter++)
{
Console.WriteLine(DateTime.Now + " running iteration " + iter);
var newWeightVector = new WeightVector(weightVector.FeatureKDictionary);
SetForwardBackwordAlgo(weightVector);
//for (var k = 0; k < weightVector.FeatureKDictionary.Count; k++)
for (var k = weightVector.FeatureKDictionary.Count-1; k >= 0; k--)
{
if (k%100 == 0)
{
Console.WriteLine(DateTime.Now + " running iteration for k " + k);
}
var wk = Compute(k, weightVector);
wk = weightVector.Get(k) + _lambda*wk;
newWeightVector.SetKey(k, wk);
}
weightVector = newWeightVector;
}
_weightVector = weightVector;
return weightVector;
}
public void ReMappingFromWeightVector(WeightVector weightVector)
{
var newDictKtoF = new Dictionary<int, string>();
var newDictFtoK = new Dictionary<string, int>();
var weightDict = new Dictionary<int, double>();
int k = 0;
var sortedDictionary = from pair in weightVector.WDictionary
where Math.Abs(pair.Value) > 2
orderby Math.Abs(pair.Value) descending
select pair;
foreach (var weight in sortedDictionary)
{
var feature = DictKToFeatures[weight.Key];
newDictFtoK[feature] = k;
newDictKtoF[k] = feature;
weightDict[k] = weight.Value;
k++;
}
weightVector.WDictionary = weightDict;
DictFeaturesToK = weightVector.FeatureKDictionary = newDictFtoK;
DictKToFeatures = newDictKtoF;
}
public void ReMapFeatureToK(bool normalize = true)
{
MapFeatures.ReMappingFromWeightVector(_useAvg ? AvgWeightVector : WeightVector, normalize);
if (_useAvg)
{
WeightVector = AvgWeightVector;
}
}
public double ComputeFunctionValue(WeightVector weightVector)
{
double outputDouble = 0;
for (var lineIndex = 0; lineIndex < _inputSentence.Count; lineIndex++)
{
var outputTags = _outputTagsList[lineIndex];
double initOutputDouble = 0;
for (var k = 0; k < weightVector.FeatureCount; k++)
{
initOutputDouble += GetAllFeatureKFromCacheWithWeights(outputTags, k,
lineIndex, weightVector);
}
initOutputDouble -= Math.Log(forwardBackwordAlgos[lineIndex].Z);
outputDouble += initOutputDouble;
}
Console.WriteLine(DateTime.Now+": new function value is: "+outputDouble);
return -outputDouble;
}
public void ComputeGradientMultiThread(WeightVector weightVector, double[] gradient,
int threadCount)
{
if (threadCount > 1)
{
//var newWeightVector = weightVector.DeepCopy();
var doneEvents = new ManualResetEvent[threadCount];
var partition = weightVector.FeatureCount / threadCount;
for (int threadIndex = 0; threadIndex < threadCount; threadIndex++)
{
var start = threadIndex*partition;
var end = start + partition;
end = end > weightVector.FeatureCount ? weightVector.FeatureCount : end;
doneEvents[threadIndex] = new ManualResetEvent(false);
var info = new ThreadInfoObject(this, start, end, weightVector,
doneEvents[threadIndex], gradient);
ThreadPool.QueueUserWorkItem(info.StartLBFGGradientComputing, threadIndex);
}
WaitHandle.WaitAll(doneEvents);
}
ComputeGradientValues(weightVector, gradient, 0, weightVector.FeatureCount);
}
public void ComputeGradientValues(WeightVector weightVector, double[] gradient,
int startIndex, int endIndex )
{
for (var k = startIndex; k < endIndex; k++)
{
double outputDouble = 0;
var kstring = "@#" + k.ToString(CultureInfo.InvariantCulture);
for (var lineIndex = 0; lineIndex < _inputSentence.Count; lineIndex++)
{
var outputTags = _outputTagsList[lineIndex];
double initOutputDouble = 0;
initOutputDouble += GetAllFeatureKFromCache(outputTags, k, lineIndex);
initOutputDouble -= CalculateGradient(outputTags, k,
lineIndex, kstring);
outputDouble += initOutputDouble;
}
gradient[k] = outputDouble - (_lambda * weightVector.Get(k));
}
}
public WeightedFeatureSum(WeightVector weightVector, List<string> sentence, bool crf = false)
{
_sentence = sentence;
_crf = crf;
WeightVector = weightVector;
}
private void SetForwardBackwordAlgo(WeightVector weightVector)
{
if (_inputSentence.Count != _outputTagsList.Count)
{
throw new Exception("counts dont match " + _inputSentence.Count + "with " + _outputTagsList.Count);
}
int counter = 0;
forwardBackwordAlgos.Clear();
foreach (var sentence in _inputSentence)
{
var outputTags = _outputTagsList[counter];
if (sentence.Count != outputTags.Count)
{
throw new Exception("counts dont match " + sentence.Count + "with " + outputTags.Count);
}
forwardBackwordAlgos.Add(new ForwardBackwordAlgo(sentence, weightVector, outputTags));
counter++;
}
}
public double GetAllFeatureKFromCacheWithWeights(List<string> tags, int k,
int lineIndex, WeightVector weightVector)
{
double sum = 0;
for (var pos = 0; pos < tags.Count; pos++)
{
var prevTag = "*";
if (pos > 0)
{
prevTag = tags[pos - 1];
}
if (_cache.Contains(prevTag, tags[pos], k, pos, lineIndex))
{
//var val = Math.Exp(_weightVector.Get(k));
sum += (_weightVector.Get(k));
}
}
return sum;
}
private double Compute(int k, WeightVector weightVector)
{
double output = 0;
//double secondTerm = 0;
int lineIndex = 0;
//var weightedFeaturesum = new WeightedFeatureSum(weightVector, null, true);
if (_inputSentence.Count != _outputTagsList.Count)
{
throw new Exception("counts dont match "+ _inputSentence.Count + "with "+ _outputTagsList.Count);
}
var ngramTags = new Tags(_tagList);
// first term.
foreach (var sentence in _inputSentence)
{
var outputTags = _outputTagsList[lineIndex];
if (sentence.Count != outputTags.Count)
{
throw new Exception("compute counts dont match " + sentence.Count + "with " + outputTags.Count);
}
output += CalculateGradient(outputTags, k,
ngramTags, lineIndex);
//output += weightedFeaturesum.GetAllFeatureK(outputTags, k, sentence);
//// second term.
//for (var j = 0; j < outputTags.Count; j++)
//{
// double sum = 0;
// foreach (var ngramTag in ngramTags.GetNGramTags(2))
// {
// string[] split = ngramTag.Split(new[] {':'});
// sum += (forwardBackwordAlgos[i].GetQ(j, split[0], split[1]) *
// weightedFeaturesum.GetFeatureK(split[0], split[1], j, k, sentence));
// }
// secondTerm += sum;
//}
lineIndex++;
}
output = output - (_lambda*weightVector.Get(k));
return output;
}
public WeightVector RunIterations(WeightVector weightVector, int iterationCount, int threadCount = 1)
{
_weightVector = weightVector;
for (var iter = 0; iter < iterationCount; iter++)
{
Console.WriteLine(DateTime.Now + " running iteration " + iter);
var newWeightVector = _weightVector.DeepCopy();
SetForwardBackwordAlgo(newWeightVector);
if (threadCount > 1)
{
var doneEvents = new ManualResetEvent[threadCount];
var partition = newWeightVector.FeatureCount / threadCount;
for (int threadIndex = 0; threadIndex < threadCount; threadIndex++)
{
var start = threadIndex*partition;
var end = start + partition;
end = end > newWeightVector.FeatureCount ? newWeightVector.FeatureCount : end;
doneEvents[threadIndex] = new ManualResetEvent(false);
var info = new ThreadInfoObject(this, start, end, newWeightVector,
doneEvents[threadIndex], null);
ThreadPool.QueueUserWorkItem(info.StartGradientComputing, threadIndex);
}
WaitHandle.WaitAll(doneEvents);
}
else
{
ComputeRange(0, _weightVector.FeatureCount, newWeightVector);
}
_weightVector = newWeightVector;
if (iter + 1 < iterationCount)
{
_weightVector.AvgNormalize();
}
}
_weightVector.AvgNormalize();
return _weightVector;
}
public void AddWeightVector(WeightVector weightVector)
{
for (int i = 0; i < weightVector.WeightArray.Length; i++)
{
WeightArray[i] += weightVector.WeightArray[i];
}
}
public void RunLBFGAlgo(WeightVector weightVector)
{
double epsg = 0.0000000001;
double epsf = 0;
double epsx = 0;
int maxits = 50;
alglib.minlbfgsstate state;
alglib.minlbfgsreport rep;
_weightVector = weightVector;
this.FeatureKDictionary = weightVector.FeatureKDictionary;
alglib.minlbfgscreate(5, weightVector.WeightArray, out state);
alglib.minlbfgssetcond(state, epsg, epsf, epsx, maxits);
alglib.minlbfgsoptimize(state, GetFunctionValueAndGradient, null, this);
double[] output;
alglib.minlbfgsresults(state, out output, out rep);
weightVector.WeightArray = output;
System.Console.WriteLine(DateTime.Now+": terminationtype {0}", rep.terminationtype); // EXPECTED: 4
//System.Console.WriteLine("{0}", alglib.ap.format(x, 2)); // EXPECTED: [-3,3]
//System.Console.ReadLine();
}
private void SetForwardBackwordAlgo(WeightVector weightVector)
{
if (_inputSentence.Count != _outputTagsList.Count)
{
throw new Exception("counts dont match " + _inputSentence.Count + "with " + _outputTagsList.Count);
}
int counter = 0;
forwardBackwordAlgos.Clear();
foreach (var sentence in _inputSentence)
{
if (counter % 100 == 0)
Console.WriteLine(DateTime.Now + "running fw/backword iteration: "+counter);
var algo = new ForwardBackwordAlgo(sentence, weightVector, _tagList);
algo.Run();
forwardBackwordAlgos.Add(algo);
counter++;
}
}
