public override void ExecuteForward(IContext context)
{
var input = context.Data;
var tensor = input.GetMatrix().ReshapeAs4DTensor(input.Rows, input.Columns, input.Depth);
var inputWidth = tensor.ColumnCount;
var inputHeight = tensor.RowCount;
var newWidth = ((inputWidth - _filterWidth + (2 * _padding)) / _xStride) + 1;
var newHeight = ((inputHeight - _filterHeight + (2 * _padding)) / _yStride) + 1;
if (_padding > 0)
{
tensor = tensor.AddPadding(_padding);
}
var im2Col = tensor.Im2Col(_filterWidth, _filterHeight, _xStride, _yStride);
var outputSignal = im2Col.Multiply(_filter);
outputSignal.AddToEachRow(_bias);
var outputTensor = outputSignal.ReshapeAs4DTensor(newHeight, newWidth);
Debug.Assert(outputTensor.Depth == FilterCount && outputTensor.Count == tensor.Count);
var graphData = new Tensor4DGraphData(outputTensor);
_AddNextGraphAction(context, graphData,
() => new Backpropagation(this, im2Col, inputWidth, inputHeight, tensor.Depth, tensor.Count,
newWidth, newHeight));
}
public override void ExecuteForward(IContext context)
{
var input = context.Data;
var lap = context.LinearAlgebraProvider;
var tensor = input.GetMatrix().ConvertTo4DTensor(input.Rows, input.Columns, input.Depth);
var inputWidth = tensor.ColumnCount;
var inputHeight = tensor.RowCount;
var newWidth = ((inputWidth - _filterWidth + (2 * _padding)) / _stride) + 1;
var newHeight = ((inputHeight - _filterHeight + (2 * _padding)) / _stride) + 1;
if (_padding > 0)
{
tensor = tensor.AddPadding(_padding);
}
var im2Col = tensor.Im2Col(_filterWidth, _filterHeight, _stride);
var outputSignal = im2Col.Multiply(_filter);
outputSignal.AddToEachRow(_bias);
var outputTensor = outputSignal.ConvertTo4DTensor(newHeight, newWidth);
var graphData = new Tensor4DGraphData(outputTensor);
_AddNextGraphAction(context, graphData, () => new Backpropagation(this, im2Col, inputWidth, inputHeight, newWidth, newHeight));
}
public override void ExecuteForward(IContext context)
{
var input = context.Data;
var tensor = input.GetMatrix().ReshapeAs4DTensor(input.Rows, input.Columns, input.Depth);
var(output, index) = tensor.MaxPool(_filterWidth, _filterHeight, _xStride, _yStride, true);
//#if DEBUG
// Debug.Assert(output.ReshapeAsVector().IsEntirelyFinite());
// Debug.Assert(index.ReshapeAsVector().IsEntirelyFinite());
//#endif
var graphData = new Tensor4DGraphData(output);
_AddNextGraphAction(context, graphData, () => new Backpropagation(this, index, tensor.ColumnCount, tensor.RowCount, output.ColumnCount, output.RowCount, output.Depth, _filterWidth, _filterHeight, _xStride, _yStride));
}
public override IMiniBatch Get(IExecutionContext executionContext, IReadOnlyList <int> rows)
{
var data = _GetRows(rows)
.Select(r => (_dataColumnIndex.Select(i => ((FloatTensor)r.Data[i]).GetAsRaw()).ToList(), ((FloatVector)r.Data[_dataTargetIndex]).Data))
.ToList()
;
var inputList = new List <IGraphData>();
for (var i = 0; i < _dataColumnIndex.Length; i++)
{
var i1 = i;
var input = _lap.CreateMatrix(InputSize, data.Count, (x, y) => data[y].Item1[i1][x]);
var tensor = new Tensor4DGraphData(input, Height, Width, Depth);
inputList.Add(tensor);
}
var output = OutputSize > 0 ? _lap.CreateMatrix(data.Count, OutputSize, (x, y) => data[x].Item2[y]) : null;
return(new MiniBatch(rows, this, inputList, new MatrixGraphData(output)));
}
