/// <summary>
/// don't use
/// </summary>
/// <returns></returns>
private double[,] LoadTopicWords()
{
var wordId = 0;
StatusMessage.Write("Loading topic words...");
var writer = StatusMessage.StatusWriter(() => string.Format("Processed {0:D0} words.", wordId));
StatusMessage.Add(writer);
var topicWords = new double[this.NumTopics, this.NumWords];
var lines =
File.ReadLines(this.wordTopicAllocationsFileName)
.Skip(TopicWordsAllocationHeaderLines)
.Take(this.NumWords);
foreach (var line in lines)
{
var parts = line
.Split(' ')
.Skip(1)
.Take(this.NumTopics)
.Select(double.Parse)
.ToArray();
for (var topic = 0; topic < this.NumTopics; topic++)
{
topicWords[topic, wordId] = parts[topic];
}
wordId++;
}
StatusMessage.Remove(writer);
return(topicWords);
}
/// <summary>
/// Process documents topic allocations file to gather:
/// * allocations per topic
/// * most prominent allocations per topic
/// * number of most prominent documents per topic
/// * top 10 documents for each topic
/// * identified bad topic
/// </summary>
/// <param name="topicsInfoFile">file to store topics info</param>
/// <returns>true if succeeded</returns>
private bool BuildTopicsInfo(string topicsInfoFile)
{
var docIndex = 0;
StatusMessage.Write("Aggregating topic document allocations...");
var writer = StatusMessage.StatusWriter(() => string.Format("Processed {0:D0} documents.", docIndex));
StatusMessage.Add(writer);
this.topicsInfo = new TopicInfo[this.NumTopics].Init();
// Differentiate if the document-topic allocations file was generated by training OR DvGen (and then copy).
// the one generated by training has the number of columns equal to NumOfTopics, whereas the one generated by DvGen has fixed number of columns = 4.
string sourceOfDocumentTopicAllocationsFile;
var lines =
File.ReadLines(this.documentTopicAllocationsFile);
if (lines.Count() < this.numDocs)
{
throw new Exception("The number of lines in document-topic allocations file is less than the number of documents!");
}
string[] firstLine = lines.First().Split(' ');
float number;
if (firstLine.Length == NumOfColumnsForDvGenOutput &&
!Single.TryParse(firstLine[1], out number) &&
!Single.TryParse(firstLine[2], out number))
{
sourceOfDocumentTopicAllocationsFile = "DvGen";
}
else if (firstLine.Length >= NumTopics &&
Single.TryParse(firstLine[0], out number) &&
Single.TryParse(firstLine[1], out number))
{
sourceOfDocumentTopicAllocationsFile = "Training";
lines = lines.Skip(this.numDocs * (this.numPasses - 1));
}
else
{
throw new Exception("Document-topic allocations file is not in right format!");
}
foreach (var line in lines)
{
// check if line is empty (this happens when a document does not contain any word in the corpus vocabulary).
if (string.IsNullOrWhiteSpace(line))
{
docIndex++;
continue;
}
IEnumerable <float> parts;
if (sourceOfDocumentTopicAllocationsFile == "Training")
{
parts = line.Split(' ')
.Take(this.NumTopics)
.Select(Single.Parse);
}
else
{
var denseVector = LoadDocumentVector(line);
parts = denseVector.ToArray();
}
var sum = parts.Sum();
var allocations = parts.Select(v => v / sum).ToList();
double max = -1;
List <int> maxTopics = new List <int>();
for (var topic = 0; topic < this.NumTopics; topic++)
{
var alloc = allocations[topic];
this.topicsInfo[topic].AggregatedAllocation += alloc;
if (alloc >= max)
{
if (alloc > max)
{
max = alloc;
maxTopics.Clear();
}
maxTopics.Add(topic);
}
}
foreach (var maxTopic in maxTopics)
{
this.topicsInfo[maxTopic].ProminentAllocation += max;
this.topicsInfo[maxTopic].ProminentFrequency++;
this.topicsInfo[maxTopic].TopProminentDocuments.Add(new Tuple <int, double>(docIndex, max));
}
docIndex++;
}
foreach (var topicInfo in this.topicsInfo)
{
topicInfo.TopProminentDocuments = topicInfo.TopProminentDocuments.OrderByDescending(x => x.Item2).Take(20).ToList();
}
var sortedTopics = Enumerable.Range(0, this.NumTopics)
.OrderByDescending(topic => this.topicsInfo[topic].AggregatedAllocation).ToArray();
int badTopicCount = 0;
var lastAllocation = this.topicsInfo[sortedTopics.Last()].AggregatedAllocation;
for (int i = sortedTopics.Length - 1; i >= 0; i--)
{
if (Math.Abs(this.topicsInfo[sortedTopics[i]].AggregatedAllocation - lastAllocation) > MaxDiffOfAllocationBetweenBadTopics)
{
break;
}
this.topicsInfo[sortedTopics[i]].IsBadTopic = true;
badTopicCount++;
}
// assume no bad topic
if (badTopicCount == 1)
{
this.topicsInfo[sortedTopics.Last()].IsBadTopic = false;
}
using (var sw = new StreamWriter(topicsInfoFile))
{
sw.WriteLine(JsonConvert.SerializeObject(this.topicsInfo, Formatting.Indented));
}
StatusMessage.Remove(writer);
return(true);
}