private IEnumerable <SuspiciousNode> SelectSuspiciousElementsWithMaskForFastRejectionLearning(
List <List <SuspiciousNode> > candidateNodesLists, BigInteger xor, BigInteger mask,
Classifier classifier)
{
for (int i = 0; i < candidateNodesLists.Count; i++)
{
var vector = BigInteger.Zero;
var classifierUnit = classifier.Units[i];
var candidates = candidateNodesLists[i]
.OrderBy(
t => LearningExperimentUtil.CountBits(
classifierUnit.Accepting
& ((t.Vector & _rejectingFeatureBitMask) ^ _rejectingFeatureBitMask)))
.ToList();
foreach (var candidate in candidates)
{
var newVector = (vector | (candidate.Vector ^ xor)) & mask;
if (newVector == vector)
{
continue;
}
vector = newVector;
yield return(candidate);
}
}
}
public IEnumerable <SuspiciousNode> SelectSuspiciousAcceptedNodes(
IReadOnlyList <List <SuspiciousNode> > candidateNodesLists,
Classifier classifier)
{
for (int i = 0; i < candidateNodesLists.Count; i++)
{
var candidates = candidateNodesLists[i];
foreach (var candidate in candidates)
{
candidate.BitsCount =
LearningExperimentUtil.CountBits(candidate.Vector & _acceptingFeatureBitMask);
}
candidates.Sort((t1, t2) => t1.BitsCount.CompareTo(t2.BitsCount));
var count = DetermineCount(classifier, i);
foreach (var candidate in candidates)
{
if (candidate.Used)
{
continue;
}
yield return(candidate);
candidate.Used = true;
if (--count == 0)
{
break;
}
}
}
}
private IEnumerable <SuspiciousNode> SelectSuspiciousElementsWithMaskWithSmallGrowing(
IReadOnlyList <List <SuspiciousNode> > candidateNodesLists, BigInteger xor, BigInteger mask)
{
for (int i = 0; i < candidateNodesLists.Count; i++)
{
var vector = BigInteger.Zero;
var candidates = candidateNodesLists[i];
while (true)
{
var minDiffCount = int.MaxValue;
SuspiciousNode newNode = null;
foreach (var candidate in candidates)
{
var newVector = (vector | (candidate.Vector ^ xor)) & mask;
var diff = newVector ^ vector;
var diffCount = LearningExperimentUtil.CountBits(diff);
if (diffCount > 0 && minDiffCount > diffCount)
{
minDiffCount = diffCount;
vector = newVector;
newNode = candidate;
}
}
if (newNode != null)
{
yield return(newNode);
}
else
{
break;
}
}
}
}
public IEnumerable <SuspiciousNode> SelectNodesForFastRejectionLearning(
IReadOnlyList <List <SuspiciousNode> > candidateNodesLists,
Classifier classifier)
{
for (int i = 0; i < candidateNodesLists.Count; i++)
{
var candidates = candidateNodesLists[i];
var rejectingVector = classifier.Units[i].Rejecting;
foreach (var cnadidate in candidates)
{
cnadidate.BitsCount = LearningExperimentUtil.CountBits(
(cnadidate.Vector & _rejectingFeatureBitMask) | rejectingVector);
}
candidates.Sort((t1, t2) => t1.BitsCount.CompareTo(t2.BitsCount));
var count = DetermineCount(classifier, i);
for (int j = candidates.Count - 1; j >= 0; j--)
{
var candidate = candidates[j];
if (candidate.Used)
{
continue;
}
yield return(candidate);
candidate.Used = true;
if (--count == 0)
{
break;
}
}
}
}
public IEnumerable <SuspiciousNode> SelectNodesForSlowAcceptanceLearning(
IReadOnlyList <List <SuspiciousNode> > candidateNodesLists,
Classifier classifier)
{
for (int i = 0; i < candidateNodesLists.Count; i++)
{
var candidates = candidateNodesLists[i];
var acceptingVector = classifier.Units[i].Accepting;
foreach (var target in candidates)
{
target.BitsCount = LearningExperimentUtil.CountBits(target.Vector & acceptingVector);
}
candidates.Sort((t1, t2) => t1.BitsCount.CompareTo(t2.BitsCount));
var count = DetermineCount(classifier, i);
for (int j = candidates.Count - 1; j >= 0; j--)
{
var candidate = candidates[j];
if (candidate.Used)
{
continue;
}
yield return(candidate);
candidate.Used = true;
if (--count == 0)
{
break;
}
}
}
}
/// <summary>
/// Update SurroundingRange, TargetRange, and Node properties then return the last index of the code processed.
/// </summary>
/// <param name="structuredCode">The structured code processed</param>
/// <param name="cst">The concrete syntax tree</param>
/// <param name="fragments"></param>
/// <returns>The updated last index of the code processed</returns>
public static List <SeedNode> ConstructAcceptingFragments(StructuredCode structuredCode, CstNode cst, IList <SelectedFragment> fragments)
{
var seedNodes = CreateSeedNodes(structuredCode, cst, fragments);
var uppermostSeedAcceptedNodes = seedNodes
.Select(node => node.Node.AncestorWithSingleChild())
.ToImmutableHashSet();
// We can select multiple nodes in corresponding to a fragment selected by a user
// and it means that we have multiple choices for selecting node names to filter nodes
// This code tries to select good node names to not filter nodes wanted by a user
var selectedNodeNames = LearningExperimentUtil.FindGoodNodeNames(uppermostSeedAcceptedNodes)
.ToImmutableHashSet();
foreach (var seedNode in seedNodes)
{
// Update the node in corresponding to the selected node names keeping the code range of the node
seedNode.Node = seedNode.Node.DescendantsOfSingleAndSelf()
.First(e => selectedNodeNames.Contains(e.Name));
var rootNode = seedNode.SurroundingRange.FindInnermostNode(cst);
var node = seedNode.Node;
var path = node.Name;
while ((node = node.Parent) != rootNode)
{
path = path + "<" + node.Name + node.RuleId;
}
seedNode.Path = path;
}
return(seedNodes);
}
public EncodingResult Encode(
ICollection <string> codePaths, IEnumerable <CstNode> allCsts,
LearningExperiment oracle, SeedNodeSet seedNodeSet = null)
{
var fileName = codePaths.Count > 0
? string.Join(",", codePaths).GetHashCode() + "_" +
(codePaths.First() + "," + codePaths.Last() + ",").GetHashCode() + "_"
+ codePaths.Count + ".encoded"
: null;
var formatter = new BinaryFormatter();
if (fileName != null && File.Exists(fileName))
{
using (var fs = new FileStream(fileName, FileMode.Open, FileAccess.Read)) {
try {
var ret = ((EncodingResult)formatter.Deserialize(fs)).MakeImmutable();
Console.WriteLine("############### Warning ###############");
Console.WriteLine("Cache file of encoded result is used.");
Console.WriteLine("#######################################");
return(ret);
} catch (Exception e) {
Console.Error.WriteLine(e);
}
}
}
var allUppermostNodes = allCsts.SelectMany(
cst => LearningExperimentUtil.GetUppermostNodesByNames(cst, _selectedNodeNames));
var result = new EncodingResult();
if (seedNodeSet != null)
{
result.SeedAcceptedNodeCount = seedNodeSet.AcceptedNodes.Count;
result.SeedNodeCount = result.SeedAcceptedNodeCount
+ seedNodeSet.RejectedNodes.Count;
EncodeSeedNodes(
seedNodeSet.AcceptedNodes, result, result.IdealAcceptedVector2GroupPath,
result.SeedAcceptedVector2GroupPath, oracle);
EncodeSeedNodes(
seedNodeSet.RejectedNodes, result, result.IdealRejectedVector2GroupPath,
result.SeedRejectedVector2GroupPath, oracle);
}
EncodeTargetNodes(allUppermostNodes, result, oracle);
if (fileName != null)
{
using (var fs = new FileStream(fileName, FileMode.Create, FileAccess.Write)) {
formatter.Serialize(fs, result);
}
}
return(result.MakeImmutable());
}
private void UpdateVector2GroupPath(
IDictionary <BigInteger, string> vector2GroupPath, BigInteger vector, CstNode node)
{
var groupPath = GetGroupPathFromNode(node);
var existingGroupPath = vector2GroupPath.GetValueOrDefault(vector);
if (existingGroupPath == null)
{
vector2GroupPath.Add(vector, groupPath);
}
else
{
vector2GroupPath[vector] = LearningExperimentUtil.GetCommonSuffix(
existingGroupPath, groupPath);
}
}
public IEnumerable <SuspiciousNode> SelectNodesForSlowRejectionLearningStrongly(
IReadOnlyList <List <SuspiciousNode> > candidateNodesLists,
Classifier classifier)
{
for (int i = 0; i < candidateNodesLists.Count; i++)
{
var candidates = candidateNodesLists[i];
var rejectingUnit = classifier.Units[i].Rejecting;
foreach (var target in candidates)
{
target.BitsCount = LearningExperimentUtil.CountBits(
(target.Vector & _rejectingFeatureBitMask) | rejectingUnit);
}
candidates.Sort((t1, t2) => t1.BitsCount.CompareTo(t2.BitsCount));
var vector = _rejectingFeatureBitMask;
var count = DetermineStrongCount(i, classifier);
foreach (var candidate in candidates)
{
if (!candidate.Used)
{
continue;
}
var newVector = vector & candidate.Vector;
if (newVector == vector)
{
continue;
}
vector = newVector;
yield return(candidate);
candidate.Used = true;
if (--count == 0)
{
break;
}
}
}
}
private SuspiciousNode SelectMostDifferentElement(
IEnumerable <BigInteger> existings, IEnumerable <SuspiciousNode> candidates,
BigInteger mask)
{
if (!existings.Any())
{
return(candidates.FirstOrDefault());
}
var maxDiff = 0;
SuspiciousNode ret = null;
foreach (var candidate in candidates)
{
var vector = candidate.Vector & mask;
var diff = existings.Min(f => LearningExperimentUtil.CountBits((f & mask) ^ vector));
if (maxDiff < diff)
{
maxDiff = diff;
ret = candidate;
}
}
return(ret);
}
public IEnumerable <SuspiciousNode> SelectNodesForFastAcceptanceLearningStrongly(
IReadOnlyList <List <SuspiciousNode> > candidateNodesLists,
Classifier classifier)
{
for (int i = 0; i < candidateNodesLists.Count; i++)
{
var candidates = candidateNodesLists[i];
var acceptingVector = classifier.Units[i].Accepting;
foreach (var target in candidates)
{
target.BitsCount = LearningExperimentUtil.CountBits(target.Vector & acceptingVector);
}
candidates.Sort((t1, t2) => t1.BitsCount.CompareTo(t2.BitsCount));
var vector = BigInteger.Zero;
var count = DetermineStrongCount(i, classifier);
foreach (var candidate in candidates)
{
if (candidate.Used)
{
continue;
}
var newVector = (vector | candidate.Vector) & _acceptingFeatureBitMask;
if (newVector == vector)
{
continue;
}
vector = newVector;
yield return(candidate);
candidate.Used = true;
if (--count == 0)
{
break;
}
}
}
}
public LearningResult Learn(
ICollection <string> seedPaths, ICollection <string> codePaths, string searchPattern,
StreamWriter writer = null)
{
var allCsts = GenerateValidCsts(codePaths);
var seedCsts = GenerateValidCsts(seedPaths).ToList();
var seedNodes = seedCsts
.SelectMany(
cst => LearningExperimentUtil.GetUppermostNodesByNames(cst, OracleNames))
.Where(ProtectedIsAcceptedUsingOracle)
.ToList();
var seedCst = seedCsts.First();
var seedCode = seedCst.Code;
var structuredCode = new StructuredCode(seedCode);
var acceptingFragments = ConstructAcceptingFragments(structuredCode, seedCst, seedNodes);
var rejectingFragments = ConstructRejectingFragments(structuredCode, seedCst);
SeedNodeSet.Create(acceptingFragments, this);
var preparingTime = Environment.TickCount;
var extractor = CreateExtractor();
var seedNodeSet = new SeedNodeSet(acceptingFragments.Select(f => f.Node), seedCsts, this);
Console.WriteLine("#Accepted seed nodes: " + seedNodeSet.AcceptedNodes.Count
+ " (" + acceptingFragments.Count + ")");
Console.WriteLine("#Rejected seed nodes: " + seedNodeSet.RejectedNodes.Count
+ " (" + rejectingFragments.Count + ")");
var featureSet = new FeatuerSet(seedNodeSet, extractor, acceptingFragments, rejectingFragments);
var groupPaths = seedNodeSet.SelectedNodeNames.Select(n => ">" + n + ">");
var classifier = new Classifier(groupPaths, featureSet);
Console.WriteLine(
"#Features: " + featureSet.AcceptingFeatureCount + ", "
+ featureSet.RejectingFeatureCount);
Console.WriteLine("Inner: " + extractor.IsInner);
var featureEncoder = new FeatureEncoder(seedNodeSet.SelectedNodeNames, extractor,
featureSet);
var encodingResult = featureEncoder.Encode(codePaths, allCsts, this, seedNodeSet);
Console.WriteLine("#Unique Elements: " + encodingResult.VectorCount);
if (encodingResult.IdealAcceptedVector2GroupPath.Keys.ToHashSet()
.Overlaps(encodingResult.IdealRejectedVector2GroupPath.Keys.ToHashSet()))
{
var others = encodingResult.IdealRejectedVector2GroupPath;
var vector =
encodingResult.IdealAcceptedVector2GroupPath.Keys.First(others.ContainsKey);
foreach (var featureString in featureEncoder.GetFeatureStringsByVector(vector))
{
Console.WriteLine(Experiment.Beautify(featureString));
}
throw new Exception("Master predicates can't classify elements!");
}
var groupCache = new GroupCache(encodingResult, classifier);
var trainingSet = encodingResult.CreateTrainingVectorSet();
classifier.Create(trainingSet, groupCache);
Experiment.WriteFeatureStrings(Console.Out, classifier, featureEncoder);
Console.WriteLine("Preparing time: " + (Environment.TickCount - preparingTime));
var count = 0;
var sumTime = Environment.TickCount;
ClassificationResult classificationResult;
while (true)
{
var time = Environment.TickCount;
classificationResult = Classify(count, classifier, groupCache, encodingResult,
trainingSet);
if (classificationResult.SuspiciousNodes == null)
{
break;
}
var additionalAcceptedSet = RevealSuspiciousElements(
encodingResult.IdealAcceptedVector2GroupPath.Keys,
classificationResult.SuspiciousNodes, encodingResult, trainingSet);
if (!classifier.Update(additionalAcceptedSet, trainingSet, groupCache))
{
count++;
}
else
{
count = 0;
}
Console.WriteLine("Time: " + (Environment.TickCount - time));
}
classifier.MakeImmutable();
Console.WriteLine();
Console.WriteLine("Sum time: " + (Environment.TickCount - sumTime));
var trainingVectorCount = trainingSet.Count;
var idealVectorCount = encodingResult.IdealVectorSet.Count;
Console.WriteLine("#Required vectors: " + trainingVectorCount + " / " + idealVectorCount);
if (writer != null)
{
encodingResult.WriteResult(writer, trainingSet);
}
foreach (var groupPath in classifier.GroupPaths)
{
Console.WriteLine(groupPath);
}
classifier.Optimize(encodingResult.IdealRejectedVector2GroupPath.Keys, groupCache);
return(new LearningResult {
ClassificationResult = classificationResult,
Classifier = classifier,
EncodingResult = encodingResult,
FeatureEncoder = featureEncoder,
});
}
public int CountUsingOracle(CstNode cst)
{
return(LearningExperimentUtil.GetUppermostNodesByNames(cst, OracleNames)
.Count(IsAcceptedUsingOracle));
}
public IEnumerable <int> CountRejectingFeatures()
{
return(Units.Select(c =>
LearningExperimentUtil.CountBits(c.Rejecting >> AcceptingFeatureCount)));
}
public IEnumerable <int> CountAcceptingFeatures()
{
return(Units.Select(c =>
LearningExperimentUtil.CountBits(c.Accepting & AcceptingFeatureBitMask)));
}
