/// <summary>
/// Converts a dataset to a set of training examples in 2D for use in CNTK.
/// </summary>
/// <typeparam name="T">Supported by<seealso cref="float"/>, <seealso cref="double"/></typeparam>
/// <param name="features">2D Feature Set</param>
/// <param name="labels">Set of labels. The dimension of the labels should be the same.</param>
/// <param name="minibatchSize">Minipack size</param>
/// <returns></returns>
public IEnumerable <Minibatch> ConvertDatasetToMinibatch <T>(IEnumerable <T[, ]> features, IEnumerable <T[]> labels, int minibatchSize) where T : IConvertible
{
var combined = features.Zip(labels, (f, l) => (f, l));
foreach (var segment in GetSegments(combined, minibatchSize))
{
var featuresData = segment.SelectMany(p => MatrixToVector(p.f));
var labelsData = segment.SelectMany(p => p.l);
Minibatch minibatch = new Minibatch();
minibatch.Size = segment.Count;
minibatch.Features = Value.CreateBatch(new int[] { GetRowsCount(segment[0].f), GetColumnsCount(segment[0].f), 1 }, featuresData, Device);
minibatch.Labels = Value.CreateBatch(new int[] { segment[0].l.Length }, labelsData, Device);
yield return(minibatch);
}
}
/// <summary>
/// Converts a dataset into sets of training examples for use in CNTK.
/// </summary>
/// <typeparam name="T">Supported by<seealso cref="float"/>, <seealso cref="double"/></typeparam>
/// <param name="dataset">Dataset. Each example should contain signs at the beginning of the array with dimensions of inputDim, and at the end of class labels with dimensions of outputDim.
/// For example, inputDim = 3, outputDim = 2: [f1, f2, f3, l1, l2]</param>
/// <param name="minibatchSize">Minipack size</param>
/// <param name="inputDim">Dimension of signs (capacity)</param>
/// <returns></returns>
public IEnumerable <Minibatch> ConvertDatasetToMinibatch <T>(IEnumerable <T[]> dataset, int inputDim, int minibatchSize) where T : IConvertible
{
var outputDim = (dataset.FirstOrDefault()?.Length ?? 0) - inputDim;
foreach (var minibatchData in GetSegments(dataset, minibatchSize))
{
var features = minibatchData.SelectMany(p => p.Take(inputDim)).ToArray();
var labels = minibatchData.SelectMany(p => p.Skip(inputDim)).ToArray();
Minibatch minibatch = new Minibatch();
minibatch.Size = minibatchData.Count;
minibatch.Features = Value.CreateBatch(new int[] { inputDim }, features, Device);
minibatch.Labels = Value.CreateBatch(new int[] { outputDim }, labels, Device);
yield return(minibatch);
}
}
/// <summary>
/// Converts a dataset into sets of training examples for use in recursive networks.
/// </summary>
/// <typeparam name="T">Supported by<seealso cref="float"/>, <seealso cref="double"/></typeparam>
/// <param name="features">A set of sequences (traits). Each sequence can be of variable length, but of the same dimension (the arrays of which the sequence consists must have the same length)</param>
/// <param name="labels">Set of labels. The dimension of the labels should be the same.</param>
/// <param name="minibatchSize">Minipack size</param>
/// <returns></returns>
public IEnumerable <Minibatch> ConvertDatasetToMinibatch <T>(IEnumerable <IList <T[]> > features, IEnumerable <T[]> labels, int minibatchSize) where T : IConvertible
{
var inputDimension = features.FirstOrDefault()?[0].Length ?? 0;
var outputDimension = labels.FirstOrDefault()?.Length ?? 0;
var combined = features.Zip(labels, (f, l) => (f, l));
foreach (var batchData in GetSegments(combined, minibatchSize))
{
//{}-miniBatch, ()-sequence, []-features.
//{ ([inputDim], [inputDim], [inputDim]), ([inputDim], [inputDim]) } => { [inputDim * 3], [inputDim * 2] }
var featuresTransformed = batchData.Select(p => p.f.SelectMany(q => q));
//{ [outputDim], [outputDim] } => { outputDim * 2 }
var labelTransformed = batchData.SelectMany(p => p.l);
Minibatch minibatch = new Minibatch();
minibatch.Features = Value.CreateBatchOfSequences(new[] { inputDimension }, featuresTransformed, Device);
minibatch.Labels = Value.CreateBatch(new[] { outputDimension }, labelTransformed, Device);
minibatch.Size = batchData.Count;
yield return(minibatch);
}
}
