/// <summary>
/// Create a random subsample of size nSamples, without replacement.
/// </summary>
/// <param name="nSamples"></param>
/// <param name="rangen"></param>
/// <returns></returns>
public QueryCollection Sample(int nSamples, Random rangen)
{
if (nSamples > nQueries)
throw new Exception("nSamples cannot exceed total number of queries.");
int[] indices = new int[nQueries];
ArrayUtils.Range(indices, 0, 1);
float[] vals = new float[nQueries];
ArrayUtils.Random(vals, 0, 1, rangen);
Array.Sort(vals, indices);
Query[] qs = new Query[nSamples];
for (int i = 0; i < nSamples; ++i)
{
qs[i] = queries[indices[i]];
}
return new QueryCollection(qs, true, 0.0);
}
/// <summary>
/// query1 and query2 must be the same query (but with different scores). The urls must be in the same order. However they need not be sorted.
/// </summary>
/// <param name="query1"></param>
/// <param name="query2"></param>
/// <param name="dcg"></param>
/// <param name="truncLevel"></param>
/// <returns>Null if this query has zero maxDCG. Else, a rankedItem array, sorted by the scores in query1.</returns>
public static RankedItem[] FillRankedItems(Query query1, Query query2, DCGScorer scorer, Random ran)
{
if(query1.Length != query2.Length)
throw new Exception("Query length mismatch.");
if(query1.QueryID != query2.QueryID)
throw new Exception("Queries have differnt IDs.");
int length = query1.Length;
double maxDCG = query1.MaxNonTruncDCG;
if(maxDCG == 0.0)
return null;
RankedItem[] rankedItems = new RankedItem[length];
double[] scores1 = query1.scores;
double[] scores2 = query2.scores;
string QID = query1.QueryID;
for(int i = 0; i < length; ++i)
{
float label = query1.Labels[i];
if(label != query2.Labels[i])
throw new Exception("FillRankedItems: label mismatch.");
rankedItems[i] = new RankedItem((double)DCGScorer.scoresMap[(int)label] / maxDCG, scores1[i], scores2[i], label);//, QueryID);
}
if (rankedItems != null)
{
SortNJitter(rankedItems, ran);
}
return rankedItems;
}
/// <summary>
/// Labels are hard copied. Feature vectors are left unallocated.
/// Space is allocated for a single column of scores.
/// </summary>
/// <param name="qc"></param>
/// <returns></returns>
public QueryCollection CopyEmptyQueryCollection()
{
Query[] newQueries = new Query[nQueries];
for (int i = 0; i < nQueries; ++i)
{
Query q = queries[i];
float[][] ftrVectors = ArrayUtils.FloatMatrix(0, 0);
double[] scores = new double[q.Length];
float[] newLabels = ArrayUtils.Copy(q.Labels);
newQueries[i] = new Query(q, newLabels, ftrVectors, scores);
}
return new QueryCollection(newQueries, this.skipDegenerateQueries, scoreForDegenerateQueries);
}
/// <summary>
/// Hard copy.
/// </summary>
/// <returns></returns>
public QueryCollection CopyQueryCollection(bool skipDegenerateQueries, double scoreForDegenerateQueries)
{
Query[] newQueries = new Query[nQueries];
for (int i = 0; i < nQueries; ++i)
{
Query q = queries[i];
float[][] ftrVectors = ArrayUtils.Copy(q.FtrVectors);
double[] scores = ArrayUtils.Copy(q.scores);
float[] newLabels = ArrayUtils.Copy(q.Labels);
newQueries[i] = new Query(q, newLabels, ftrVectors, scores);
}
return new QueryCollection(newQueries, skipDegenerateQueries, this.scoreForDegenerateQueries);
}
/// <summary>
/// Generate a random query collection for testing. The scores are correlated with the labels, and the labels are distributed
/// according to priors (priors[0] = P(Bad), etc.).
/// </summary>
/// <param name="nQueries"></param>
/// <param name="priors"></param>
/// <param name="scale"></param>
/// <param name="nScores"></param>
/// <param name="nRowsPerQuery"></param>
/// <param name="rangen"></param>
public QueryCollection(int nQueries, float[] priors, double scale, int nScores, int nRowsPerQuery, Random rangen)
{
this.nQueries = nQueries;
nRows = nRowsPerQuery * nQueries;
queries = new Query[nQueries];
for (int i = 0; i < nQueries; ++i)
{
queries[i] = new Query(priors, scale, nScores, nRowsPerQuery, i.ToString(), rangen);
}
}
/// <summary>
/// Assumes all query data has been prepared in the linked list of QueryRows, except features and scores (which don't live
/// in individual queries).
/// </summary>
/// <param name="QID"></param>
/// <param name="dcg">Can be null if desired (in which case max DCGs won't be computed).</param>
public Query(string QID, double scoreForDegenerateQuery)
{
this.QueryID = QID;
this.scoreForDegenerateQuery = scoreForDegenerateQuery;
this.length = 0;
QueryRow ptr = QueryRow.mostRecent;
while(ptr != null)
{
++length;
ptr = ptr.previous;
}
ftrVectors = new float[length][];
scores = new double[length];
scoresCp = new double[length];
ranks = new int[length];
labels = new float[length];
ptr = QueryRow.mostRecent;
int ctr = length-1; // stick to original order
while(ptr != null)
{
labels[ctr] = ptr.Label;
ftrVectors[ctr] = ptr.Features;
--ctr;
ptr = ptr.previous;
}
// Reset QueryRow
QueryRow.Reset();
previous = mostRecent;
mostRecent = this;
// Fill DCGs
DCGScorer dcg = new DCGScorer();
FillMaxDCGs(dcg);
}
/// <summary>
/// A degenerate query is one with all the labels the same.
/// </summary>
/// <param name="fname"></param>
/// <param name="skipDegenerate"></param>
public QueryCollection(string fname, float labelForUnlabeled, bool skipDegenerateQueries, double scoreForDegenerateQuery)
{
this.skipDegenerateQueries = skipDegenerateQueries;
this.scoreForDegenerateQueries = scoreForDegenerateQuery;
using(StreamReader sr = new StreamReader(fname))
{
string[] headers = sr.ReadLine().Split(QueryRowHeader.Separator);
QueryRowHeader queryRowHeader = new QueryRowHeader(headers);
#if USE_BM25_TO_BREAK_DEGENERACY
degenBreak_idx = queryRowHeader.DegenBreak_idx;
#endif
nRows = 0;
string row = sr.ReadLine();
string[] splitRow;
string lastQID = null;
string QID = null;
Query.Reset();
while(row != null)
{
splitRow = row.Split(QueryRowHeader.Separator);
QID = splitRow[queryRowHeader.queryIDIdx];
if(QID != lastQID && nRows != 0)
{
new Query(lastQID, scoreForDegenerateQuery);
++nQueries;
}
string rating = splitRow[queryRowHeader.ratingIdx];
float label;
switch(rating)
{
case "Definitive":
label = 4;
break;
case "Perfect":
label = 4;
break;
case "Excellent":
label = 3;
break;
case "Good":
label = 2;
break;
case "Fair":
label = 1;
break;
case "Bad":
label = 0;
break;
case "Detrimental":
label = 0;
break;
case "Unknown":
label = -1;
break;
case "": // Unlabeled (in RatedNRandom)
label = -1;
break;
case "HighlyRelevant": // ImageSearch - happily the only label that overlaps with MSNSearch("Detrimental") gets the same score
label = 2;
break;
case "Relevant":
label = 1;
break;
case "NotRelevant":
label = 0;
break;
case "Unjudged":
label = 0;
break;
default:
try
{
label = float.Parse(rating);
}
catch(Exception)
{
Console.WriteLine("Unable to parse rating " + rating + " into an float. Using 0");
label = 0;
}
break;
}
// Convention is: if row has shorter length than number of headers, the missing values are all zero.
//float[] ftrVector = new float[queryRowHeader.FeatureCount];
float[] ftrVector = new float[queryRowHeader.FeatureCount];
//for(int j = 0, i = queryRowHeader.firstFtrIdx; j < queryRowHeader.FeatureCount; ++i, ++j)
for (int j = 0, i = queryRowHeader.firstFtrIdx; i < splitRow.Length; ++i, ++j)
{
string val = splitRow[i];
ftrVector[j] = ( val == String.Empty ) ? 0.0F : float.Parse(val);
}
new QueryRow(label, ftrVector, labelForUnlabeled);
lastQID = QID;
row = sr.ReadLine();
++nRows;
}
new Query(QID, scoreForDegenerateQuery);
++nQueries;
}
// Finally construct the array, in the original order
queries = new Query[nQueries];
Query ptr = Query.mostRecent;
int ctr = nQueries-1;
while(ptr != null)
{
queries[ctr--] = ptr;
ptr = ptr.previous;
}
// No need to fix unlabeled rows, since QueryRow does that. (Any unlabeled rows must be fixed before degenerates are removed).
FixUnlabeledRows(labelForUnlabeled);
// For training, for efficiency, we might want to skip queries that have no pairs.
if (skipDegenerateQueries)
{
int nonDegCtr = 0;
int newNRows = 0;
for (int i = 0; i < nQueries; ++i)
{
Query q = queries[i];
if (!q.AllRowsSameLabel())
{
queries[nonDegCtr++] = q;
newNRows += q.Length;
}
}
if (nonDegCtr < nQueries)
{
Query[] newQueries = new Query[nonDegCtr]; // wish C# had direct way to shorten array
Array.Copy(queries, newQueries, nonDegCtr);
nRows = newNRows;
nQueries = nonDegCtr;
queries = newQueries;
}
}
Query.Reset(); // Prepare for next load
}
public QueryCollection(Query[] queries, bool skipDegenerateQueries, double scoreForDegenerateQueries)
{
this.skipDegenerateQueries = skipDegenerateQueries;
this.scoreForDegenerateQueries = scoreForDegenerateQueries;
this.queries = queries;
nQueries = queries.Length;
nRows = 0;
for(int i = 0; i < nQueries; ++i)
{
nRows += queries[i].Length;
}
}
internal static void Reset()
{
mostRecent = null;
}
/// <summary>
/// If maxDCG and maxTruncDCG have been passed, it's safe to assume that the labels have already been remapped,
/// hence labelForUnlabeled is not needed.
/// </summary>
/// <param name="QID"></param>
/// <param name="labels"></param>
/// <param name="scores"></param>
/// <param name="maxDCG"></param>
/// <param name="maxTruncDCG"></param>
public Query(Query q, float[] labels, float[][] ftrVectors, double[] scores)
{
if (ftrVectors.Length != 0 && labels.Length != ftrVectors.Length) // Allow for zero-sized feature vectors (for queries that contain only scores)
throw new Exception("Query constructor: size mismatch");
QueryID = q.QueryID;
this.scoreForDegenerateQuery = q.scoreForDegenerateQuery;
this.maxNonTruncDCG = q.maxNonTruncDCG;
this.maxTruncDCG = q.maxTruncDCG;
this.labels = labels;
this.ftrVectors = ftrVectors;
this.scores = scores;
this.maxNonTruncDCG = q.maxNonTruncDCG;
this.maxTruncDCG = q.maxTruncDCG;
length = labels.Length;
}
private QueryCollection CreateQueryCollection()
{
int cQueris = this.labelFeatureDataCoded.DataGroups.GroupCounts;
Query[] queries = new Query[cQueris];
for (int qIdx = 0; qIdx < cQueris; qIdx++)
{
DataGroup queryGroup = this.labelFeatureDataCoded.DataGroups[qIdx];
Query query = CreateQuery(queryGroup, this.labels, this.score, this.labelForUnlabeled, this.scoreForDegenerateQuery);
queries[qIdx] = query;
}
bool skipDegenerateQueries = true;
QueryCollection qc = new QueryCollection(queries, skipDegenerateQueries, this.scoreForDegenerateQuery);
return qc;
}
private Query CreateQuery(DataGroup queryGroup, float[] inLabels, float[] inScores,
float labelForUnlabeled, double scoreForDegenerateQuery)
{
string QID = queryGroup.id.ToString();
float[] labels = new float[queryGroup.cSize];
double[] scores = new double[queryGroup.cSize];
int end = queryGroup.iStart + queryGroup.cSize;
for (int i = queryGroup.iStart; i < end; i++)
{
labels[i - queryGroup.iStart] = (float)inLabels[i];
scores[i - queryGroup.iStart] = inScores[i];
}
DCGScorer.truncLevel = this.truncLevel;
Query query = new Query(QID, labels, null, scores, labelForUnlabeled, scoreForDegenerateQuery);
return query;
}
private float AbsDeltaNDCG(RankPair rankPair, DataGroup queryGroup, Query query)
{
int idx1 = rankPair.IdxL - queryGroup.iStart;
int idx2 = rankPair.IdxH - queryGroup.iStart;
return query.AbsDeltaNDCG(idx1, idx2);
}