public IReadOnlyList <IRecurrentOutput> Execute(IReadOnlyList <float[]> inputData)
{
var context = new List <IDisposableMatrixExecutionLine>();
context.Add(new DisposableMatrixExecutionLine());
context.Add(new DisposableMatrixExecutionLine());
var ret = new List <RecurrentOutput>();
using (var m2 = _initialMemory.ToRowMatrix()) {
context[1].Assign(m2);
foreach (var item in inputData)
{
using (var curr = _lap.Create(item))
using (var curr2 = curr.ToRowMatrix()) {
context[0].Assign(curr2);
foreach (var action in _layer)
{
action.Activate(context);
}
var memoryOutput = context[1].Current.AsIndexable().Rows.First();
var output = context[0].Current.Row(0).AsIndexable();
ret.Add(new RecurrentOutput(output, memoryOutput));
}
}
}
return(ret);
}
public LogisticRegression GradientDescent(int iterations, float learningRate, float lambda = 0.1f, Func <float, bool> costCallback = null)
{
var theta = _lap.Create(_feature.ColumnCount, 0f);
for (var i = 0; i < iterations; i++)
{
if (costCallback != null)
{
var cost = ComputeCost(theta, lambda);
if (!costCallback(cost))
{
break;
}
}
using (var d = _Derivative(theta, lambda)) {
d.Multiply(learningRate);
var theta2 = theta.Subtract(d);
theta.Dispose();
theta = theta2;
}
}
var ret = new LogisticRegression {
Theta = theta.Data
};
theta.Dispose();
return(ret);
}
public IMiniBatch GetTrainingData(IReadOnlyList <int> rows)
{
var input = _lap.Create(rows.Count, _inputSize, (x, y) => Get(rows[x], y));
var output = _lap.Create(rows.Count, _outputSize, (x, y) => GetPrediction(rows[x], y));
return(new MiniBatch(input, output));
}
public IMiniBatch GetTrainingData(IReadOnlyList <int> rows)
{
var batchRow = _table.GetRows(rows).Select(r => _Convert(r)).ToList();
var input = _lap.Create(batchRow.Count, _inputSize, (x, y) => batchRow[x].Item1[y]);
var output = _lap.Create(batchRow.Count, _outputSize, (x, y) => batchRow[x].Item2[y]);
return(new MiniBatch(input, output));
}
public Standard(ILinearAlgebraProvider lap, int inputSize, int outputSize, LayerDescriptor init, IActivationFunction activation, IWeightInitialisation weightInit)
{
_descriptor = init;
_activation = activation;
// initialise weights and bias
_bias = lap.Create(outputSize, x => weightInit.GetBias());
_weight = lap.Create(inputSize, outputSize, (x, y) => weightInit.GetWeight(inputSize, outputSize, x, y));
}
public void TestRandomProjection()
{
var a = _lap.Create(256, 256, (x, y) => x * y).AsIndexable();
var projector = _lap.CreateRandomProjection(256, 32);
var projections = projector.Compute(a);
Assert.IsTrue(projections.ColumnCount == 32);
Assert.IsTrue(projections.RowCount == 256);
}
protected void _CreateFilter(IMatrix matrix)
{
Debug.Assert(_filter == null);
// create a row level probability
//var dropout = Enumerable.Range(0, matrix.ColumnCount).Select(v => _probabilityDistribution.Sample() / _invertedMultiplier).ToArray();
// create a filter against the dropout probability
_filter = _lap.Create(matrix.RowCount, matrix.ColumnCount, (x, y) => _probabilityDistribution.Sample() / _invertedMultiplier);
}
public RegressionTrainer(ILinearAlgebraProvider lap, IDataTable table)
{
_lap = lap;
var numRows = table.RowCount;
var numCols = table.ColumnCount;
int classColumnIndex = table.TargetColumnIndex;
var data = table.GetNumericColumns(Enumerable.Range(0, numCols).Where(c => c != classColumnIndex));
_feature = lap.Create(numRows, numCols, (i, j) => j == 0 ? 1 : data[j - 1][i]);
_target = lap.Create(table.GetColumn <float>(classColumnIndex));
}
public ISequentialMiniBatch GetTrainingData(int sequenceLength, IReadOnlyList <int> rows)
{
var input = new IMatrix[sequenceLength];
var output = new IMatrix[sequenceLength];
var dataGroup = _inputData[sequenceLength];
for (var k = 0; k < sequenceLength; k++)
{
input[k] = _lap.Create(rows.Count, _inputSize, (x, y) => dataGroup[rows[x]].Item1[k].Input[y]);
output[k] = _lap.Create(rows.Count, _outputSize, (x, y) => dataGroup[rows[x]].Item1[k].Output[y]);
}
return(new SequentialMiniBatch(input, output, rows.Select(r => dataGroup[r].Item2).ToArray()));
}
public InternalLayer(ILinearAlgebraProvider lap, int inputSize, int outputSize, IActivationFunction activation, ConvolutionDescriptor descriptor, bool disableUpdate)
{
_inputSize = inputSize;
_outputSize = outputSize;
_activation = activation;
_descriptor = descriptor;
_disableUpdate = disableUpdate;
var weightInit = lap.NN.GetWeightInitialisation(descriptor.WeightInitialisation);
_bias = lap.Create(outputSize, x => weightInit.GetBias());
_weight = lap.Create(inputSize, outputSize, (x, y) => weightInit.GetWeight(inputSize, outputSize, x, y));
}
StandardFeedForward _ReadFeedForward(NetworkLayer layer)
{
var descriptor = LayerDescriptor.CreateFrom(layer);
var bias = _lap.Create(layer.OutputSize, 0f);
bias.Data = layer.Bias;
var weight = _lap.Create(layer.InputSize, layer.OutputSize, 0f);
weight.Data = layer.Weight;
return(new StandardFeedForward(weight, bias, _activation[descriptor.Activation]));
}
public INeuralNetworkRecurrentBackpropagation Execute(List <IMatrix> curr, bool backpropagate)
{
var input = curr[0];
var memory = curr[1];
var a = Combine(input, memory, _wc.Layer, _uc.Layer, m => _activation.Calculate(m));
var i = Combine(input, memory, _wi.Layer, _ui.Layer, m => m.SigmoidActivation());
var f = Combine(input, memory, _wf.Layer, _uf.Layer, m => m.SigmoidActivation());
var o = Combine(input, memory, _wo.Layer, _uo.Layer, m => m.SigmoidActivation());
using (var f2 = f.PointwiseMultiply(memory)) {
var ct = a.PointwiseMultiply(i);
ct.AddInPlace(f2);
var cta = _activation.Calculate(ct);
curr[0] = o.PointwiseMultiply(cta);
curr[1] = ct;
if (backpropagate)
{
var ones = _lap.Create(memory.RowCount, memory.ColumnCount, (x, y) => 1f);
return(new Backpropagation(_activation, ones, ct, cta, memory, o, a, i, f, input, _uc, _wc, _ui, _wi, _uf, _wf, _uo, _wo));
}
//memory.Dispose();
//input.Dispose();
a.Dispose();
i.Dispose();
f.Dispose();
o.Dispose();
cta.Dispose();
return(null);
}
}
public NNMF(ILinearAlgebraProvider lap, IReadOnlyList <IIndexableVector> data, int numClusters, IErrorMetric costFunction = null)
{
_lap = lap;
_data = data;
_numClusters = numClusters;
_costFunction = costFunction ?? ErrorMetricType.RMSE.Create();
// create the main matrix
var rand = new Random();
_dataMatrix = _lap.Create(data.Count, data.First().Count, (x, y) => data[x][y]);
// create the weights and features
_weights = _lap.Create(_dataMatrix.RowCount, _numClusters, (x, y) => Convert.ToSingle(rand.NextDouble()));
_features = _lap.Create(_numClusters, _dataMatrix.ColumnCount, (x, y) => Convert.ToSingle(rand.NextDouble()));
}
public void TiedAutoEncoder()
{
const int DATA_SIZE = 1000, REDUCED_SIZE = 200;
// create some random data
var rand = new Random();
var trainingData = _lap.NN.CreateTrainingDataProvider(Enumerable.Range(0, 100)
.Select(i => _lap.Create(DATA_SIZE, v => Convert.ToSingle(rand.NextDouble())))
.Select(v => new TrainingExample(v.Data.Data, v.Data.Data))
.ToList()
);
var layerTemplate = new LayerDescriptor(0f)
{
Activation = ActivationType.Relu,
WeightUpdate = WeightUpdateType.RMSprop
};
var firstLayer = _lap.NN.CreateLayer(DATA_SIZE, REDUCED_SIZE, layerTemplate);
var secondLayer = _lap.NN.CreateTiedLayer(firstLayer, layerTemplate);
var layers = new[] {
_lap.NN.CreateTrainer(firstLayer, layerTemplate),
_lap.NN.CreateTrainer(secondLayer, layerTemplate)
};
var errorMetric = ErrorMetricType.RMSE.Create();
using (var trainer = _lap.NN.CreateBatchTrainer(layers)) {
var trainingContext = _lap.NN.CreateTrainingContext(errorMetric, 0.03f, 32);
trainer.Train(trainingData, 2, trainingContext);
}
}
internal PerWeightUpdateBase(INeuralNetworkLayerUpdater layerUpdater, ILinearAlgebraProvider lap)
{
_layerUpdater = layerUpdater;
var targetWeight = layerUpdater.Layer.Weight;
_cache = lap.Create(targetWeight.RowCount, targetWeight.ColumnCount, (x, y) => 0f);
}
public I3DTensor ExecuteToTensor(I3DTensor tensor)
{
using (var output = ExecuteToMatrix(tensor)) {
// convert the matrix to a tensor
var sliceList = new List <IMatrix>();
for (int i = 0, len = output.ColumnCount; i < len; i++)
{
using (var vector = output.Column(i)) {
var parts = vector.Split(tensor.ColumnCount);
var sliceMatrix = _lap.Create(parts);
sliceList.Add(sliceMatrix);
foreach (var part in parts)
{
part.Dispose();
}
}
}
var ret = _lap.CreateTensor(sliceList);
foreach (var slice in sliceList)
{
slice.Dispose();
}
return(ret);
}
}
public IMatrix Execute(IMatrix error, ITrainingContext context, bool calculateOutput, INeuralNetworkUpdateAccumulator updateAccumulator)
{
var matrixList = error.AsIndexable().Columns.Select(v => v.ToArray()).ToList();
var newMatrixList = new List <IMatrix>();
Tuple <int, int> newIndex;
for (var i = 0; i < matrixList.Count; i++)
{
var matrix = matrixList[i];
var table = _indexPosList[i];
newMatrixList.Add(_lap.Create(_rows, _columns, (x, y) => {
if (table.TryGetValue(Tuple.Create(x, y), out newIndex))
{
var newIndex2 = newIndex.Item1 * _newRows + newIndex.Item2;
return(matrix[newIndex2]);
}
return(0f);
}));
}
using (var tensor = _lap.CreateTensor(newMatrixList)) {
var ret = tensor.ConvertToMatrix();
foreach (var item in newMatrixList)
{
item.Dispose();
}
return(ret);
}
}
public I3DTensor ConvertToTensor(IMatrix matrix)
{
var sliceList = new List <IMatrix>();
for (int i = 0, len = matrix.ColumnCount; i < len; i++)
{
using (var vector = matrix.Column(i)) {
var parts = vector.Split(_inputWidth);
//var sliceMatrix = _lap.Create(parts).Transpose();
var sliceMatrix = _lap.Create(parts);
sliceList.Add(sliceMatrix);
foreach (var part in parts)
{
part.Dispose();
}
}
}
var ret = _lap.CreateTensor(sliceList);
foreach (var slice in sliceList)
{
slice.Dispose();
}
return(ret);
}
public AdamUpdater(INeuralNetworkLayerUpdater layerUpdater, ILinearAlgebraProvider lap, float decay, float decay2) : base(layerUpdater, lap)
{
_decay = decay;
_decay2 = decay2;
var targetWeight = layerUpdater.Layer.Weight;
_cache2 = lap.Create(targetWeight.RowCount, targetWeight.ColumnCount, (x, y) => 0f);
}
public IMiniBatch GetTrainingData(IReadOnlyList <int> rows)
{
_backpropagation.Clear();
var rowList = new List <IVector>();
var outputList = new List <float[]>();
foreach (var item in rows)
{
var data = _data[item];
var tensor = data.Item1;
var backpropagation = _isTraining ? new Stack <IConvolutionalLayerBackpropagation>() : null;
for (int i = 0, len = _layer.Count; i < len - 1; i++)
{
var next = _layer[i].ExecuteToTensor(tensor, backpropagation);
if (tensor != data.Item1)
{
tensor.Dispose();
}
tensor = next;
}
rowList.Add(_layer.Last().ExecuteToVector(tensor, backpropagation));
if (tensor != data.Item1)
{
tensor.Dispose();
}
if (backpropagation != null)
{
_backpropagation.Add(backpropagation);
}
outputList.Add(data.Item2);
}
var input = _lap.Create(rowList);
foreach (var item in rowList)
{
item.Dispose();
}
var output = _lap.Create(rows.Count, _outputSize, (x, y) => outputList[x][y]);
return(new MiniBatch(input, output));
}
public TiedLayer(ILinearAlgebraProvider lap, INeuralNetworkLayer layer, IWeightInitialisation weightInit)
{
_inputSize = layer.OutputSize;
_outputSize = layer.InputSize;
_layer = layer;
_weight = layer.Weight;
_bias = lap.Create(_outputSize, x => weightInit.GetBias());
_weightTranspose = _weight.Transpose();
}
public float[] Predict(IReadOnlyList <IReadOnlyList <float> > input)
{
using (var feature = _lap.Create(input.Count, input[0].Count + 1, (i, j) => j == 0 ? 1 : input[i][j - 1]))
using (var h0 = feature.Multiply(_theta))
using (var h1 = h0.Column(0))
using (var h = h1.Sigmoid())
using (var h2 = h.AsIndexable()) {
return(h2.ToArray());
}
}
public RandomProjection(ILinearAlgebraProvider lap, int fixedSize, int reducedSize, int s = 3)
{
_lap = lap;
_fixedSize = fixedSize;
_reducedSize = reducedSize;
var c1 = Math.Sqrt(3);
var distribution = new Categorical(new[] { 1.0 / (2 * s), 1 - (1.0 / s), 1.0 / (2 * s) });
_matrix = _lap.Create(fixedSize, reducedSize, (i, j) => Convert.ToSingle((distribution.Sample() - 1) * c1));
}
public KNNClassifier(ILinearAlgebraProvider lap, KNearestNeighbours model, int k, DistanceMetric distanceMetric = DistanceMetric.Euclidean)
{
_k = k;
_lap = lap;
_model = model;
_distanceMetric = distanceMetric;
for (int i = 0, len = model.Instance.Length; i < len; i++)
{
_instance.Add(lap.Create(model.Instance[i].Data));
}
}
// normal method removed until GPU provider can properly calculate matrix inverses!
//public LinearRegression Solve()
//{
// // solve using normal method
// using (var lambdaMatrix = _lap.CreateIdentity(_feature.ColumnCount))
// using (var zero = _lap.Create(1, 0f)) {
// lambdaMatrix.UpdateColumn(0, zero.AsIndexable(), 0);
// using (var featureTranspose = _feature.Transpose())
// using (var pinv = featureTranspose.Multiply(_feature))
// using (var pinv2 = pinv.Add(lambdaMatrix))
// using (var pinv3 = pinv2.Inverse())
// using (var tc = _target.ToColumnMatrix())
// using (var a2 = featureTranspose.Multiply(tc))
// using (var ret = pinv3.Multiply(a2))
// using (var theta = ret.Column(0)) {
// return new LinearRegression {
// Theta = theta.Data
// };
// }
// }
//}
public LinearRegression GradientDescent(int iterations, float learningRate, float lambda = 0.1f, Func <float, bool> costCallback = null)
{
var regularisation = 1f - (learningRate * lambda) / _feature.RowCount;
var theta = _lap.Create(_feature.ColumnCount, 0f);
using (var regularisationVector = _lap.Create(theta.Count, i => i == 0 ? 1f : regularisation)) {
for (var i = 0; i < iterations; i++)
{
if (costCallback != null)
{
var cost = ComputeCost(theta, lambda);
if (!costCallback(cost))
{
break;
}
}
using (var p = _feature.Multiply(theta))
using (var pc = p.Column(0))
using (var e = pc.Subtract(_target))
using (var e2 = e.ToRowMatrix())
using (var d = e2.Multiply(_feature))
using (var delta = d.Row(0)) {
delta.Multiply(learningRate);
using (var temp = theta.PointwiseMultiply(regularisationVector)) {
var theta2 = temp.Subtract(delta);
theta.Dispose();
theta = theta2;
}
}
}
}
var ret = new LinearRegression {
Theta = theta.Data
};
theta.Dispose();
return(ret);
}
public IReadOnlyList <IVector> GetNumericColumns(ILinearAlgebraProvider lap, IEnumerable <int> columns = null)
{
var columnTable = (columns ?? Enumerable.Range(0, ColumnCount)).ToDictionary(i => i, i => new float[RowCount]);
int index = 0;
_Iterate(row => {
foreach (var item in columnTable)
{
item.Value[index] = row.GetField <float>(item.Key);
}
++index;
return(true);
});
return(columnTable.OrderBy(kv => kv.Key).Select(kv => lap.Create(kv.Value)).ToList());
}
public IReadOnlyList <IVector> GetNumericRows(ILinearAlgebraProvider lap, IEnumerable <int> columns = null)
{
var columnList = new List <int>(columns ?? Enumerable.Range(0, ColumnCount));
var ret = new List <IVector>();
_Iterate(row => {
int index = 0;
var buffer = new float[columnList.Count];
foreach (var item in columnList)
{
buffer[index++] = row.GetField <float>(item);
}
ret.Add(lap.Create(buffer));
return(true);
});
return(ret);
}
IEnumerable <Tuple <string, float> > _Classify(IRow row)
{
// encode the features into a vector
var featureCount = _model.FeatureColumn.Length;
var features = new float[featureCount];
for (var i = 0; i < featureCount; i++)
{
features[i] = row.GetField <float>(_model.FeatureColumn[i]);
}
// find the k closest neighbours and score the results based on proximity to rank the classifications
using (var vector = _lap.Create(features)) {
var distances = vector.FindDistances(_instance, _distanceMetric).AsIndexable();
return(distances.Values
.Zip(_model.Classification, (s, l) => Tuple.Create(l, s))
.OrderBy(d => d.Item2)
.Take(_k)
.GroupBy(d => d.Item1)
.Select(g => Tuple.Create(g.Key, g.Sum(d => 1f / d.Item2)))
);
}
}
public IReadOnlyList <IReadOnlyList <IVector> > Cluster(IReadOnlyList <IVector> data, int numIterations, float errorThreshold = 0.001f)
{
if (data.Count == 0)
{
return(new List <IVector[]>());
}
// create the main matrix
var data2 = new List <IIndexableVector>();
foreach (var item in data)
{
data2.Add(item.AsIndexable());
}
using (var v = _lap.Create(data.Count, data.First().Count, (x, y) => data2[x][y])) {
data2.ForEach(d => d.Dispose());
// create the weights and features
var rand = new Random();
var weights = _lap.Create(v.RowCount, _numClusters, (x, y) => Convert.ToSingle(rand.NextDouble()));
var features = _lap.Create(_numClusters, v.ColumnCount, (x, y) => Convert.ToSingle(rand.NextDouble()));
// iterate
float lastCost = 0;
for (int i = 0; i < numIterations; i++)
{
using (var wh = weights.Multiply(features)) {
var cost = _DifferenceCost(v, wh);
if (i % (numIterations / 10) == 0)
{
Console.WriteLine("NNMF cost: " + cost);
}
if (cost <= errorThreshold)
{
break;
}
lastCost = cost;
using (var wT = weights.Transpose())
using (var hn = wT.Multiply(v))
using (var wTw = wT.Multiply(weights))
using (var hd = wTw.Multiply(features))
using (var fhn = features.PointwiseMultiply(hn)) {
features.Dispose();
features = fhn.PointwiseDivide(hd);
}
using (var fT = features.Transpose())
using (var wn = v.Multiply(fT))
using (var wf = weights.Multiply(features))
using (var wd = wf.Multiply(fT))
using (var wwn = weights.PointwiseMultiply(wn)) {
weights.Dispose();
weights = wwn.PointwiseDivide(wd);
}
}
}
// weights gives cluster membership
var documentClusters = weights.AsIndexable().Rows.Select((c, i) => Tuple.Create(i, c.MaximumIndex())).ToList();
weights.Dispose();
features.Dispose();
return(documentClusters.GroupBy(d => d.Item2).Select(g => g.Select(d => data[d.Item1]).ToArray()).ToList());
}
}
/// <summary>
/// Converts the image to a matrix
/// </summary>
/// <param name="lap"></param>
/// <returns></returns>
public IMatrix AsMatrix(ILinearAlgebraProvider lap)
{
return(lap.Create(Width, Height, (i, j) => Data[j * Width + i]));
}
