public static void Serialize(DataSourceSet dss, string path, bool compress = false)
{
using (var stream = new FileStream(path, FileMode.Create, FileAccess.Write, FileShare.None))
{
Serialize(dss, stream, compress);
}
}
private static DataSourceSet ConvertFromSerializableObject(Dictionary <string, Tuple <float[], int[]> > obj)
{
var dss = new DataSourceSet();
foreach (var entry in obj)
{
dss.Add(entry.Key, DataSourceFactory.Create(entry.Value.Item1, entry.Value.Item2));
}
return(dss);
}
protected override void EndProcessing()
{
if (ParameterSetName == "load")
{
Path = IO.GetAbsolutePath(this, Path);
var dss = DataSourceSet.Load(Path, !NoDecompress);
WriteObject(dss);
}
else
{
var dss = new DataSourceSet(DataSources);
WriteObject(dss);
}
}
public static void Serialize(DataSourceSet dss, Stream stream, bool compress = false)
{
var serializer = MessagePackSerializer.Get <Dictionary <string, Tuple <float[], int[]> > >();
var obj = ConvertToSerializableObject(dss);
if (compress)
{
using (var zstream = new DeflateStream(stream, CompressionLevel.Optimal, true))
{
serializer.Pack(zstream, obj);
}
}
else
{
serializer.Pack(stream, obj);
}
}
public static IEnumerable <CTFSample> GetSampleReader(TextReader reader)
{
int lineCount = 0;
int sequenceCount = 0;
int seqStartLineCount = 1;
string line;
string seqId = null;
var splitLines = new List <string[]>();
var comments = new List <string>();
var startIndexMap = new Dictionary <string, int>();
var endIndexMap = new Dictionary <string, int>();
while ((line = reader.ReadLine()) != null)
{
++lineCount;
var splitLine = line.Split(new char[] { '|' });
var n = splitLine[0].Trim();
if (seqId == null || n == seqId)
{
splitLines.Add(splitLine);
seqId = n;
if (reader.Peek() != -1)
{
continue;
}
}
++sequenceCount;
var seqDim = splitLines.Count;
var dss = new DataSourceSet();
comments.Clear();
startIndexMap.Clear();
endIndexMap.Clear();
for (var i = 0; i < splitLines.Count; ++i)
{
var columns = splitLines[i];
for (var j = 1; j < columns.Length; ++j)
{
var feature = columns[j].Trim();
if (feature[0] == '#')
{
int skip;
for (skip = 1; skip < feature.Length && Char.IsWhiteSpace(feature[skip]); ++skip)
{
;
}
comments.Add(feature.Substring(skip));
continue;
}
var items = feature.Split();
if (items.Length < 2)
{
throw new InvalidDataException(string.Format("line {0}: Invalid feature", lineCount));
}
var featureDim = items.Length - 1;
var name = items[0];
DataSourceBase <float, float[]> ds;
float[] data;
if (dss.Features.ContainsKey(name))
{
ds = (DataSourceBase <float, float[]>)dss.Features[name];
data = ds.TypedData;
}
else
{
data = new float[featureDim * seqDim];
ds = DataSourceFactory.Create(data, new int[] { featureDim, seqDim, 1 });
dss.Add(name, ds);
startIndexMap[name] = i;
}
var baseIndex = ds.Shape.GetSequentialIndex(new int[] { 0, i, 0 });
for (var k = 0; k < featureDim; ++k)
{
data[baseIndex + k] = Converter.ToFloat(items[k + 1]);
}
endIndexMap[name] = i;
}
}
foreach (var name in dss.Features.Keys.ToArray())
{
var start = startIndexMap[name];
var end = endIndexMap[name];
if (start == 0 && end == splitLines.Count - 1)
{
continue;
}
var ds = dss[name];
dss.Features[name] = ds.Subset(start, end - start + 1, -2);
}
yield return(new CTFSample()
{
LineCount = seqStartLineCount,
SequenceCount = sequenceCount,
SequenceId = seqId,
DataSet = dss,
Comments = comments
});
splitLines.Clear();
splitLines.Add(splitLine);
seqStartLineCount = lineCount;
seqId = n;
}
}
public void AddMinibatch(DataSourceSet dataSourceSet)
{
_dataQueue.Add(dataSourceSet);
}
private static Dictionary <string, Tuple <float[], int[]> > ConvertToSerializableObject(DataSourceSet dss)
{
var obj = new Dictionary <string, Tuple <float[], int[]> >();
foreach (var entry in dss.Features)
{
var ds = entry.Value;
obj[entry.Key] = new Tuple <float[], int[]>(ds.Data.ToArray(), ds.Shape.Dimensions);
}
return(obj);
}
public static void Write(string path, DataSourceSet dataSourceSet, bool hasSequenceAxis)
{
using (var writer = new StreamWriter(path, false, new UTF8Encoding(false)))
Write(writer, dataSourceSet, hasSequenceAxis);
}
public static void Write(TextWriter writer, DataSourceSet dataSourceSet, bool withSequenceAxis)
{
var builder = new CTFBuilder(writer, 0, false);
// Argument check
var sampleCount = dataSourceSet.Features.First().Value.Shape[-1];
var maxSeqLength = 1;
foreach (var entry in dataSourceSet)
{
var name = entry.Key;
var ds = entry.Value;
if (withSequenceAxis && ds.Shape.Rank < 3)
{
throw new ArgumentException("DataSource shape should have sequence and batch axes as the last two");
}
if (!withSequenceAxis && ds.Shape.Rank < 2)
{
throw new ArgumentException("DataSource shape should have a batch axis");
}
var count = ds.Shape[-1];
if (count != sampleCount)
{
throw new ArgumentException("Sample counts of data sources should be equal");
}
if (withSequenceAxis)
{
var seqLength = ds.Shape[-2];
if (seqLength > maxSeqLength)
{
maxSeqLength = seqLength;
}
}
}
for (var sampleIndex = 0; sampleIndex < sampleCount; ++sampleIndex)
{
for (var seq = 0; seq < maxSeqLength; ++seq)
{
foreach (var entry in dataSourceSet)
{
var name = entry.Key;
var ds = entry.Value;
int seqLength;
int dim;
if (withSequenceAxis)
{
seqLength = ds.Shape[-2];
if (seq >= seqLength)
{
continue;
}
dim = ds.Shape.GetSize(ds.Shape.Rank - 3);
}
else
{
seqLength = 1;
dim = ds.Shape.GetSize(ds.Shape.Rank - 2);
}
int index = sampleIndex * dim * seqLength + seq * dim;
builder.AddDenseSample(name, new ListSlice <float>(ds.Data, index, dim));
}
builder.NextLine();
}
builder.NextSequence();
}
builder.Finish();
}
public static Value[] Invoke(this Function func, DataSourceSet dataSet, DeviceDescriptor device = null, bool errorWhenArgumentUnused = true)
{
return(Invoke(func, (IDictionary <string, IDataSource <float> >)dataSet.Features, device, errorWhenArgumentUnused));
}
