// shuffle data and write them to different partions (different files on disk for now)
public static List <DataVector> ShuffleDataAndGetInitialCentriods(string originalDataFile, int partitionsNum, int clustersNum, string executionDirectory)
{
List <DataVector> data = DataPartitionCache.ReadDataFile(originalDataFile);
// shuffle, not truely random, but sufficient for our purpose
data = data.OrderBy(a => Guid.NewGuid()).ToList();
string dataDirectory = Path.Combine(executionDirectory, Constants.DataDirectory);
// clean things up first
if (Directory.Exists(dataDirectory))
{
Directory.Delete(dataDirectory, true);
}
Directory.CreateDirectory(dataDirectory);
int residualCount = data.Count;
int batchSize = data.Count / partitionsNum;
for (int i = 0; i < partitionsNum; i++)
{
int linesCount = residualCount > batchSize ? batchSize : residualCount;
using (StreamWriter writer = new StreamWriter(
File.OpenWrite(Path.Combine(executionDirectory, Constants.DataDirectory, i.ToString(CultureInfo.InvariantCulture)))))
{
for (int j = i * batchSize; j < (i * batchSize) + linesCount; j++)
{
writer.WriteLine(data[j].ToString());
}
writer.Close();
}
}
return(InitializeCentroids(clustersNum, data, executionDirectory));
}
public LegacyKMeansTask(
DataPartitionCache dataPartition,
[Parameter(Value = typeof(KMeansConfiguratioinOptions.K))] int clustersNumber,
[Parameter(Value = typeof(KMeansConfiguratioinOptions.ExecutionDirectory))] string executionDirectory)
{
_dataPartition = dataPartition;
_clustersNum = clustersNumber;
_kMeansExecutionDirectory = executionDirectory;
if (_centroids == null)
{
string centroidFile = Path.Combine(_kMeansExecutionDirectory, Constants.CentroidsFile);
_centroids = new Centroids(DataPartitionCache.ReadDataFile(centroidFile));
}
}
public byte[] Call(byte[] memento)
{
// TODO: this belongs to dedicated data loader layer, will refactor once we have that
_groupCommClient.Initialize();
string centroidFile = Path.Combine(_kMeansExecutionDirectory, Constants.CentroidsFile);
_centroids = new Centroids(DataPartitionCache.ReadDataFile(centroidFile));
float loss = float.MaxValue;
float newLoss;
while (true)
{
if (_isInitialIteration)
{
// broadcast initial centroids to all slave nodes
Logger.Log(Level.Info, "Broadcasting INITIAL centroids to all slave nodes: " + _centroids);
_isInitialIteration = false;
}
else
{
ProcessedResults results = _meansReducerReceiver.Reduce();
_centroids = new Centroids(results.Means.Select(m => m.Mean).ToList());
Logger.Log(Level.Info, "Broadcasting new centroids to all slave nodes: " + _centroids);
newLoss = results.Loss;
Logger.Log(Level.Info, string.Format(CultureInfo.InvariantCulture, "The new loss value {0} at iteration {1} ", newLoss, _iteration));
if (newLoss > loss)
{
_controlBroadcastSender.Send(ControlMessage.STOP);
throw new InvalidOperationException(
string.Format(CultureInfo.InvariantCulture, "The new loss {0} is larger than previous loss {1}, while loss function must be monotonically decreasing across iterations", newLoss, loss));
}
else if (newLoss.Equals(loss))
{
Logger.Log(Level.Info, string.Format(CultureInfo.InvariantCulture, "KMeans clustering has converged with a loss value of {0} at iteration {1} ", newLoss, _iteration));
break;
}
else
{
loss = newLoss;
}
}
_controlBroadcastSender.Send(ControlMessage.RECEIVE);
_dataBroadcastSender.Send(_centroids);
_iteration++;
}
_controlBroadcastSender.Send(ControlMessage.STOP);
return(null);
}
public KMeansSlaveTask(
DataPartitionCache dataPartition,
[Parameter(typeof(KMeansConfiguratioinOptions.TotalNumEvaluators))] int clustersNumber,
IGroupCommClient groupCommClient)
{
using (Logger.LogFunction("KMeansSlaveTask::KMeansSlaveTask"))
{
_dataPartition = dataPartition;
_groupCommClient = groupCommClient;
_clustersNum = clustersNumber;
_commGroup = _groupCommClient.GetCommunicationGroup(Constants.KMeansCommunicationGroupName);
_dataBroadcastReceiver = _commGroup.GetBroadcastReceiver <Centroids>(Constants.CentroidsBroadcastOperatorName);
_partialMeansSender = _commGroup.GetReduceSender <ProcessedResults>(Constants.MeansReduceOperatorName);
_controlBroadcastReceiver = _commGroup.GetBroadcastReceiver <ControlMessage>(Constants.ControlMessageBroadcastOperatorName);
}
}
public byte[] CallWithWritingToFileSystem(byte[] memento)
{
string centroidFile = Path.Combine(_kMeansExecutionDirectory, Constants.CentroidsFile);
_centroids = new Centroids(DataPartitionCache.ReadDataFile(centroidFile));
_dataPartition.LabelData(_centroids);
_partialMeans = ComputePartialMeans();
// should be replaced with Group Communication
using (StreamWriter writer = new StreamWriter(
File.OpenWrite(Path.Combine(_kMeansExecutionDirectory, Constants.DataDirectory, Constants.PartialMeanFilePrefix + _dataPartition.Partition))))
{
for (int i = 0; i < _partialMeans.Count; i++)
{
writer.WriteLine(_partialMeans[i].ToString());
}
writer.Close();
}
return(null);
}
