/// <summary>
/// Run Dawid-Skene on the data.
/// </summary>
/// <param name="data">The data.</param>
/// <param name="fullData">The full data.</param>
/// <param name="calculateAccuracy">Whether to calculate accuracy</param>
/// <returns>A results instance</returns>
public Results RunDawidSkene(IList <Datum> data, IList <Datum> fullData, bool calculateAccuracy)
{
// If you want to run Dawid-Skene code, download his code, integrate it into
// the project, and change false to true below.
Console.WriteLine("--- Dawid Skene ---");
PredictedLabel = new Dictionary <string, int?>();
Mapping = new DataMapping(data);
var fullDataMapping = new DataMapping(fullData);
var labelings = data.Select(d => new Labeling(d.WorkerId, d.TaskId, d.WorkerLabel.ToString(), d.GoldLabel.ToString())).ToList();
DawidSkene ds = new DawidSkene(labelings, null, null);
// The labels may be in a different order from our data labeling - we need to create a map.
int[] labelIndexMap = new int[Mapping.LabelCount];
var dwLabels = ds.classes.Keys.ToArray();
for (int i = 0; i < Mapping.LabelCount; i++)
{
labelIndexMap[i] = Array.IndexOf(dwLabels, (i + Mapping.LabelMin).ToString());
}
GoldLabels = fullDataMapping.GetGoldLabelsPerTaskId().
ToDictionary(kvp => kvp.Key, kvp => kvp.Value == null ? (int?)null : (int?)labelIndexMap[kvp.Value.Value]);
ds.Estimate(10);
var inferredLabels = ds.GetObjectClassProbabilities().Select(r => new Discrete(r)).ToArray();
TrueLabel = inferredLabels.Select((lab, i) => new
{
key = Mapping.TaskIndexToId[i],
val = lab
}).ToDictionary(a => a.key, a => a.val);
if (calculateAccuracy)
{
UpdateAccuracy();
}
return(this);
}
/// <summary>
/// Runs the majority vote method on the data.
/// </summary>
/// <param name="data">The data</param>
/// <param name="fullData">The full data</param>
/// <param name="calculateAccuracy">Compute the accuracy (true).</param>
/// <param name="useVoteDistribution">The true label is sampled from the vote distribution (true) or it is
/// taken as the mode of the vote counts (false).
/// In the latter case, ties are broken by sampling from the most voted classes.</param>
/// <returns>The updated results</returns>
public Results RunMajorityVote(IList <Datum> data, IList <Datum> fullData, bool calculateAccuracy, bool useVoteDistribution)
{
PredictedLabel = new Dictionary <string, int?>();
Mapping = new DataMapping(data);
FullMapping = new DataMapping(fullData);
GoldLabels = FullMapping.GetGoldLabelsPerTaskId();
var inferredLabels = useVoteDistribution ? Mapping.GetVoteDistribPerTaskIndex() : Mapping.GetMajorityVotesPerTaskIndex().Select(mv => mv == null ? (Discrete)null : Discrete.PointMass(mv.Value, Mapping.LabelCount)).ToArray();
TrueLabel = inferredLabels.Select((lab, i) => new
{
key = Mapping.TaskIndexToId[i],
val = lab
}).ToDictionary(a => a.key, a => a.val);
if (calculateAccuracy)
{
UpdateAccuracy();
}
return(this);
}
/// <summary>
/// Run Dawid-Skene on the data.
/// </summary>
/// <param name="data">The data.</param>
/// <param name="fullData">The full data.</param>
/// <param name="calculateAccuracy">Whether to calculate accuracy</param>
/// <returns>A results instance</returns>
public Results RunDawidSkene(IList<Datum> data, IList<Datum> fullData, bool calculateAccuracy)
{
// If you want to run Dawid-Skene code, download his code, integrate it into
// the project, and change false to true below.
Console.WriteLine("--- Dawid Skene ---");
PredictedLabel = new Dictionary<string, int?>();
Mapping = new DataMapping(data);
var fullDataMapping = new DataMapping(fullData);
var labelings = data.Select(d => new Labeling(d.WorkerId, d.TaskId, d.WorkerLabel.ToString(), d.GoldLabel.ToString())).ToList();
DawidSkene ds = new DawidSkene(labelings, null, null);
// The labels may be in a different order from our data labeling - we need to create a map.
int[] labelIndexMap = new int[Mapping.LabelCount];
var dwLabels = ds.classes.Keys.ToArray();
for (int i = 0; i < Mapping.LabelCount; i++)
{
labelIndexMap[i] = Array.IndexOf(dwLabels, (i + Mapping.LabelMin).ToString());
}
GoldLabels = fullDataMapping.GetGoldLabelsPerTaskId().
ToDictionary(kvp => kvp.Key, kvp => kvp.Value == null ? (int?)null : (int?)labelIndexMap[kvp.Value.Value]);
ds.Estimate(10);
var inferredLabels = ds.GetObjectClassProbabilities().Select(r => new Discrete(r)).ToArray();
TrueLabel = inferredLabels.Select((lab, i) => new
{
key = Mapping.TaskIndexToId[i],
val = lab
}).ToDictionary(a => a.key, a => a.val);
if (calculateAccuracy)
{
UpdateAccuracy();
}
return this;
}
/// <summary>
/// Runs the majority vote method on the data.
/// </summary>
/// <param name="data">The data</param>
/// <param name="fullData">The full data</param>
/// <param name="calculateAccuracy">Compute the accuracy (true).</param>
/// <param name="useVoteDistribution">The true label is sampled from the vote distribution (true) or it is
/// taken as the mode of the vote counts (false).
/// In the latter case, ties are broken by sampling from the most voted classes.</param>
/// <returns>The updated results</returns>
public Results RunMajorityVote(IList<Datum> data, IList<Datum> fullData, bool calculateAccuracy, bool useVoteDistribution)
{
PredictedLabel = new Dictionary<string, int?>();
Mapping = new DataMapping(data);
FullMapping = new DataMapping(fullData);
GoldLabels = FullMapping.GetGoldLabelsPerTaskId();
var inferredLabels = useVoteDistribution ? Mapping.GetVoteDistribPerTaskIndex() : Mapping.GetMajorityVotesPerTaskIndex().Select(mv => mv == null ? (Discrete)null : Discrete.PointMass(mv.Value, Mapping.LabelCount)).ToArray();
TrueLabel = inferredLabels.Select((lab, i) => new
{
key = Mapping.TaskIndexToId[i],
val = lab
}).ToDictionary(a => a.key, a => a.val);
if (calculateAccuracy)
{
UpdateAccuracy();
}
return this;
}
