/// <summary>
/// PPT 20141202
/// 为每个source annotation求出可能的target annotations，并按照概率排序。
/// 存储到ProbabilityConstant.Pr_T_S里。
/// 注意：本函数不是训练，而是计算结果，所以遍历的是全部的sentence，不是用于train的sentence。
/// </summary>
static public void Pr_T_S()
{
//初始化
CascadedConstant.Pr_T_S = new Dictionary <SourceAnnotation, IDictionary <TargetAnnotation, double> >();
CascadedConstant.Pr_T_t = new Dictionary <TargetLabeltruth, IDictionary <TargetAnnotation, double> >();
foreach (Label label in Constant.TargetTaxonomy.LabelArray)
{
CascadedConstant.Pr_T_t.Add(new TargetLabeltruth(label, true), new Dictionary <TargetAnnotation, double>());
CascadedConstant.Pr_T_t.Add(new TargetLabeltruth(label, false), new Dictionary <TargetAnnotation, double>());
}
//开始计算
foreach (Sentence sentence in Constant.SentenceList)
{
if (!CascadedConstant.Pr_T_S.ContainsKey(sentence.GoldSourceAnnotation))
{
IDictionary <TargetAnnotation, double> targetAnnotationProbabilityDic = new Dictionary <TargetAnnotation, double>();
//遍历所有可能的target annotation
foreach (Label[] targetLabelArray in CascadedFunction.PowerSet(Constant.TargetTaxonomy.LabelArray))
{
TargetAnnotation T = new TargetAnnotation(targetLabelArray);
double probability = sentence.GoldSourceAnnotation.Pr_T_S(T); //算值
if (probability != 0) //忽略等于0的
{
targetAnnotationProbabilityDic.Add(T, probability);
}
}
//排序（已查明为什么sadness true总是最高）
IDictionary <TargetAnnotation, double> sortedElements = GeneralFunction.SortDictionary(targetAnnotationProbabilityDic);
CascadedConstant.Pr_T_S.Add(sentence.GoldSourceAnnotation, sortedElements);
}
}
}
static private void initializeEkmanAsTarget(Corpus corpus, int startIndex, int endIndex, IList <Sentence> sentences, ref IList <TargetWorker> targetWorkerList)
{
string[] data = File.ReadAllLines(corpus + "/EkmanData" + startIndex + "-" + endIndex + ".csv");
foreach (string row in data)
{
string[] labels = row.Split(',');
TargetWorker worker = new TargetWorker(labels[0]);
if (!targetWorkerList.Contains(worker))//重复的人不再添加
{
targetWorkerList.Add(worker);
}
else
{
worker = targetWorkerList.First(x => x.Equals(worker));
}
IList <Label> trueLabels = new List <Label>();
for (int i = 1; i <= (endIndex - startIndex + 1) * (Constant.EkmanLabelArray.Length + 1); ++i)
{
switch (labels[i])
{
case "Anger":
trueLabels.Add(Label.Anger);
break;
case "Sadness":
trueLabels.Add(Label.Sadness);
break;
case "Joy":
trueLabels.Add(Label.Joy);
break;
case "Disgust":
trueLabels.Add(Label.Disgust);
break;
case "Surprise":
trueLabels.Add(Label.Surprise);
break;
case "Fear":
trueLabels.Add(Label.Fear);
break;
}
if (i % (Constant.EkmanLabelArray.Length + 1) == 0)
{
//取出SentenceList里的一个Sentence
Sentence sentence = sentences[startIndex + (i - 1) / (Constant.EkmanLabelArray.Length + 1)];
TargetAnnotation targetAnnotation = new TargetAnnotation(trueLabels.ToArray());
trueLabels.Clear();
worker.SentenceTargetAnnotationDic.Add(sentence, targetAnnotation);
sentence.TargetWorkerTargetAnnotationDic.Add(worker, targetAnnotation);
}
}
}
}
static public double Similarity(TargetAnnotation a, TargetAnnotation b)
{
double result = 0;
switch (Constant.Similarity)
{
case Interoperability.Similarity.SMC:
foreach (Label label in Constant.TargetTaxonomy.LabelArray)
{
if (a.LabelAndTruthDic[label] == b.LabelAndTruthDic[label])
{
++result;
}
}
result /= Constant.TargetTaxonomy.LabelArray.Length;
break;
case Interoperability.Similarity.Jaccard:
double numerator = 0;
double denominator = 0;
foreach (Label label in Constant.TargetTaxonomy.LabelArray)
{
if (a.LabelAndTruthDic[label] && b.LabelAndTruthDic[label])
{
++numerator;
++denominator;
}
else if (a.LabelAndTruthDic[label] || b.LabelAndTruthDic[label])
{
++denominator;
}
}
result = numerator / denominator;
break;
case Interoperability.Similarity.Dice:
numerator = 0;
denominator = 0;
foreach (Label label in Constant.TargetTaxonomy.LabelArray)
{
if (a.LabelAndTruthDic[label] && b.LabelAndTruthDic[label])
{
numerator += 2;
denominator += 2;
}
else if (a.LabelAndTruthDic[label] || b.LabelAndTruthDic[label])
{
++denominator;
}
}
result = (numerator + 1) / (denominator + 1);
break;
}
return(result);
}
void Start()
{
Debug.Log(Directory.GetCurrentDirectory());
GameObject agent = transform.parent.gameObject;
annotations = agent.GetComponent <TargetAnnotation>();
initializer = agent.GetComponent <RandomInit>();
positionDrawer = agent.GetComponent <RandomPosition>();
positionDrawer.agent = transform.gameObject;
rbody = GetComponent <Rigidbody>();
engine = transform.Find("Engine").GetComponent <Engine>();
accelerometer = transform.Find("Accelerometer").GetComponent <Accelerometer>();
depthSensor = transform.Find("DepthSensor").GetComponent <DepthSensor>();
// calculate max velocity with set parameters
maxVelocity = new Vector3(engine.maxForceLateral / (rbody.drag * rbody.mass),
engine.maxForceVertical / (rbody.drag * rbody.mass),
engine.maxForceLongitudinal / (rbody.drag * rbody.mass));
maxYawVelocity = engine.maxTorqueYaw / (rbody.inertiaTensor.y * rbody.angularDrag);
}
void OnValidate()
{
GameObject agent = transform.parent.gameObject;
annotations = agent.GetComponent <TargetAnnotation>();
initializer = agent.GetComponent <RandomInit>();
positionDrawer = agent.GetComponent <RandomPosition>();
positionDrawer.agent = transform.gameObject;
if (mode == RobotAcademy.DataCollection.gate)
{
annotations.target = gateTargetObject;
positionDrawer.target = gateTargetObject;
}
else if (mode == RobotAcademy.DataCollection.path)
{
annotations.target = pathTargetObject;
positionDrawer.target = pathTargetObject;
}
}
/// <summary>
/// 计算已知Target Label为true求Target Annotation的概率。
/// </summary>
/// <param name="targetAnnotation">Target Annotation</param>
/// <returns>
/// 已知Target Label的Truth求Target Annotation的概率。
/// 如Target Label的Truth与其在Target Annotation的Truth不符，则返回0。
/// </returns>
/// （观察重复使用情况）
public double Pr_T_t(TargetAnnotation targetAnnotation)
{
#region 符时直接返回0
if (targetAnnotation.LabelAndTruthDic[this.Label] != this.Truth)
{
return(0);
}
#endregion
#region 符时求Pr_T和Pr_t
//如已经计算过，直接取即可，不用重新计算。
//因为符合t的T共有2^|T|个，所以对于每个t都有2^|T|个T对应。
if (CascadedConstant.Pr_T_t[this].ContainsKey(targetAnnotation))
{
return(CascadedConstant.Pr_T_t[this][targetAnnotation]);
}
else
{
double nT = 0; //分子
double nt = 0; //分母
foreach (Sentence sentence in TrainConstant.SentenceList)
{
foreach (TargetAnnotation targetannotation in sentence.TargetWorkerTargetAnnotationDic.Values)
{
if (targetannotation.LabelAndTruthDic[this.Label] == this.Truth)
{
++nt;
if (targetAnnotation.Equals(targetannotation))
{
++nT;
}
}
}
}
double pribability = nT / nt;
//存储，以备以后再用
CascadedConstant.Pr_T_t[this].Add(targetAnnotation, pribability);
return(pribability);
}
#endregion
}
static private void initializeNakaAsTarget(Corpus corpus, int startIndex, int endIndex, IList <Sentence> sentences, ref IList <TargetWorker> targetWorkerList)
{
string[] data = File.ReadAllLines(corpus + "/NakaData" + startIndex + "-" + endIndex + ".csv");//不需要单独做一个data-sample,因为sentence-sample以外的sentence在下面的for循环里遍历不到。
foreach (string row in data)
{
string[] labels = row.Split(','); //labels[0]是用户名
TargetWorker worker = new TargetWorker(labels[0]);
if (!targetWorkerList.Contains(worker)) //重复的人不再添加
{
targetWorkerList.Add(worker);
}
else
{
worker = targetWorkerList.First(x => x.Equals(worker));
}
IList <Label> trueLabels = new List <Label>();
for (int i = 1; i <= (endIndex - startIndex + 1) * (Constant.NakaLabelArray.Length + 1); ++i)
{
switch (labels[i])
{
case "happiness":
trueLabels.Add(Label.喜Joy);
break;
case "fondness":
trueLabels.Add(Label.好Fondness);
break;
case "relief":
trueLabels.Add(Label.安Relief);
break;
case "anger":
trueLabels.Add(Label.怒Anger);
break;
case "sadness":
trueLabels.Add(Label.哀Sadness);
break;
case "fear":
trueLabels.Add(Label.怖Fear);
break;
case "shame":
trueLabels.Add(Label.恥Shame);
break;
case "disgust":
trueLabels.Add(Label.厭Disgust);
break;
case "excitement":
trueLabels.Add(Label.昂Excitement);
break;
case "surprise":
trueLabels.Add(Label.驚Surprise);
break;
}
if (i % (Constant.NakaLabelArray.Length + 1) == 0)
{
//取出SentenceList里的一个Sentence
Sentence sentence = sentences[startIndex + (i - 1) / (Constant.NakaLabelArray.Length + 1)];
TargetAnnotation targetAnnotation = new TargetAnnotation(trueLabels.ToArray());
trueLabels.Clear();
worker.SentenceTargetAnnotationDic.Add(sentence, targetAnnotation);
sentence.TargetWorkerTargetAnnotationDic.Add(worker, targetAnnotation);
}
}
}
}
/// <summary>
/// 输出每个setnence的最优target anntoation, i.e., 概率最大的target annotation。
/// </summary>
/// <param name="method">方法</param>
/// <param name="groupsize">组容量</param>
/// <param name="groupindex">组号</param>
static public void OutputEstimatedBinaryTargetAnnotations(Method method, int groupsize, int groupindex)
{
string path = "Output/" + Constant.Gold + "/" + TrainConstant.Corpus + "_" + Constant.SourceTaxonomy.Name + "/" + method;
if (!Directory.Exists(path))
{
Directory.CreateDirectory(path);
}
StreamWriter File = new StreamWriter(path + "/" + groupsize + "-" + groupindex + method + "_BinaryTargetAnnotations.csv", false, Encoding.Default);
File.Write("Sentence,");
foreach (Label label in Constant.TargetTaxonomy.LabelArray)
{
File.Write(label + ",");
}
File.WriteLine("GashCode");
switch (method)
{
case Method.MLE:
foreach (Sentence sentence in Constant.SentenceList)
{
File.Write(sentence.ToString() + ",");
TargetAnnotation targetAnnotation = sentence.MLETargetAnnotation();
foreach (Label label in Constant.TargetTaxonomy.LabelArray)
{
File.Write((targetAnnotation.LabelAndTruthDic[label] ? 1 : 0) + ",");
}
File.WriteLine(targetAnnotation.GetHashCode());
}
break;
case Method.Cascaded:
foreach (Sentence sentence in Constant.SentenceList)
{
File.Write(sentence.ToString() + ",");
TargetAnnotation targetAnnotation = sentence.CascadedTargetAnnotation();
foreach (Label label in Constant.TargetTaxonomy.LabelArray)
{
File.Write((targetAnnotation.LabelAndTruthDic[label] ? 1 : 0) + ",");
}
File.WriteLine(targetAnnotation.GetHashCode());
}
break;
case Method.Aggregation:
case Method.OrdinaryCombination:
case Method.WeightedCombination:
case Method.ExpertiseCombination:
case Method.TemporaryNogeneralNonormalize:
case Method.TemporaryNogeneralNormalize:
case Method.TemporaryGeneralNonormalize:
case Method.TemporaryGeneralNormalize:
case Method.OtherNogeneralNormalize:
case Method.OtherGeneralNonormalize:
case Method.OtherGeneralNormalize:
foreach (Sentence sentence in Constant.SentenceList)
{
File.Write(sentence.ToString() + ",");
TargetAnnotation targetAnnotation = sentence.SpaceTargetAnnotation();
foreach (Label label in Constant.TargetTaxonomy.LabelArray)
{
File.Write((targetAnnotation.LabelAndTruthDic[label] ? 1 : 0) + ",");
}
File.WriteLine(targetAnnotation.GetHashCode());
}
break;
}
File.Close();
}
