private static void NormalizeMetrics(string metricsFile, float[] metricMeans, float[] metricStdevs)
{
StringBuilder modelSummary = new StringBuilder();
bool foundRawMetrics = false;
List <Topic> topics = new List <Topic>();
float totalAllocations = 0.0f;
long totalPromimentDF = 0L;
int numTopics = 0;
int goodTopicCount = 0;
foreach (var line in File.ReadAllLines(metricsFile))
{
if (!foundRawMetrics)
{
// Read first part.
if (!line.StartsWith("Topic Id"))
{
modelSummary.AppendLine(line.Replace("NaN", "0"));
var pairs = line.Split('\t');
if (pairs[0].ToLowerInvariant() == "topics")
{
numTopics = int.Parse(pairs[1]);
}
else if (pairs[0].ToLowerInvariant() == "good topics")
{
goodTopicCount = int.Parse(pairs[1]);
}
}
else
{
foundRawMetrics = true;
}
continue;
}
var parts = line.Split('\t');
if (parts[1] == "1")
{
// Skip bad topics.
break;
}
// Cope with wierd case that "TopDocs" column contains two double quotes.
// e.g. \\br1iceml001\ModelRepository\Models\LDA\en-us\msn\20150212_250.tsv\500_45_0.5_0.1_0.2_256_2_1_0.5\build\ExtrinsicMetrics.tsv
if (parts[7].IndexOf('\"') >= 0)
{
parts[7] = parts[7].Replace("\"", "");
}
// Read second part and reconstruct each topic
List <Tuple <int, double> > promDocs;
if (parts[6] == "0")
{
promDocs = new List <Tuple <int, double> >();
}
else
{
if (parts[7].IndexOf('|') >= 0)
{
promDocs = parts[7].Split(',')
.Select(p => new Tuple <int, double>(int.Parse(p.Split('|')[0]), double.Parse(p.Split('|')[1])))
.ToList();
}
else
{
promDocs = parts[7].Split(',')
.Select(p => new Tuple <int, double>(int.Parse(p), -1.0))
.ToList();
}
}
Topic t = new Topic()
{
TopicId = int.Parse(parts[0]),
IsBad = parts[1] == "1",
Allocations = float.Parse(parts[2]),
TC = float.Parse(parts[3]),
TS = (parts[4] == "NaN" ? 0.0f : float.Parse(parts[4])),
TD = float.Parse(parts[5]),
PromimentDF = long.Parse(parts[6]),
TopProminentDocuments = promDocs,
HighProbWords = parts.Skip(8).ToList()
};
// Compute the normalized TC/TS/TD for each topic
t.NormalizedTC = (t.TC - metricMeans[0]) / metricStdevs[0];
t.NormalizedTS = (t.TS - metricMeans[1]) / metricStdevs[1];
t.NormalizedTD = (t.TD - metricMeans[2]) / metricStdevs[2];
totalAllocations += t.Allocations;
totalPromimentDF += t.PromimentDF;
topics.Add(t);
}
// Add "Bad Topic Count" to modelSummary
int badTopicCount = topics.Count(t => t.TC < ThresholdForBadTopics);
modelSummary.AppendFormat("Bad Topic Count\t{0}\r\n", badTopicCount);
modelSummary.AppendFormat("Good Topics(%)\t{0}\r\n", (float)goodTopicCount / numTopics);
modelSummary.AppendFormat("Actual Good Topics\t{0}\r\n", goodTopicCount - badTopicCount);
// todo: add other model summary statistics here (need to modify metrics list in Model Loader as well).
DenseVector normalizedTCVector = new DenseVector(topics.Select(t => t.NormalizedTC).ToArray());
DenseVector normalizedTSVector = new DenseVector(topics.Select(t => t.NormalizedTS).ToArray());
DenseVector normalizedTDVector = new DenseVector(topics.Select(t => t.NormalizedTD).ToArray());
DenseVector weightsByAlloc = new DenseVector(topics.Select(t => t.Allocations / totalAllocations * topics.Count).ToArray());
DenseVector weightsByPromDF = new DenseVector(topics.Select(t => (float)t.PromimentDF / totalPromimentDF * topics.Count).ToArray());
double[,] cumulativeNormalizedMetricsMatrix = new double[3, 3];
// TC
cumulativeNormalizedMetricsMatrix[0, 0] = VectorBase.DotProduct(normalizedTCVector, weightsByAlloc);
cumulativeNormalizedMetricsMatrix[0, 1] = VectorBase.DotProduct(normalizedTCVector, weightsByPromDF);
cumulativeNormalizedMetricsMatrix[0, 2] = topics.Sum(t => (double)t.NormalizedTC);
// TS
cumulativeNormalizedMetricsMatrix[1, 0] = VectorBase.DotProduct(normalizedTSVector, weightsByAlloc);
cumulativeNormalizedMetricsMatrix[1, 1] = VectorBase.DotProduct(normalizedTSVector, weightsByPromDF);
cumulativeNormalizedMetricsMatrix[1, 2] = topics.Sum(t => (double)t.NormalizedTS);
// TD
cumulativeNormalizedMetricsMatrix[2, 0] = VectorBase.DotProduct(normalizedTDVector, weightsByAlloc);
cumulativeNormalizedMetricsMatrix[2, 1] = VectorBase.DotProduct(normalizedTDVector, weightsByPromDF);
cumulativeNormalizedMetricsMatrix[2, 2] = topics.Sum(t => (double)t.NormalizedTD);
// append the 3x3 matrix elements to modelSummary, save as "ExtrinsicMetrics.processed.tsv"
RegenerateModelSynopsis(metricsFile, modelSummary, topics, cumulativeNormalizedMetricsMatrix);
}
/// <summary>
/// Computes de Euclidean distance between 2 unit vectors
/// </summary>
/// <param name="v1">Unit LDA topic vector.</param>
/// <param name="v2">Unit LDA topic vector.</param>
/// <returns></returns>
public static double Distance(DenseVector v1, DenseVector v2)
{
return(Math.Sqrt((2 - 2 * VectorBase.DotProduct(v1, v2))));
}