public void ConvUtils_ReshapeConvolutionsToRowMajor()
{
var batchSize = 5;
var filterHeight = 2;
var filterWidth = 2;
var filterDepth = 2;
var stride = 1;
var padding = 0;
var inputWidth = 3;
var inputHeight = 3;
var inputDepth = 3;
var filterGridWidth = ConvUtils.GetFilterGridLength(inputWidth, filterWidth, stride, padding, BorderMode.Valid);
var filterGridHeight = ConvUtils.GetFilterGridLength(inputHeight, filterHeight, stride, padding, BorderMode.Valid);
var k = filterDepth;
var crs = inputDepth * filterWidth * filterHeight;
var npq = batchSize * filterGridWidth * filterGridHeight;
var convolutedInput = Matrix <float> .Build.Dense(k, npq, new float[] { -6.260461f, 87.38299f, -7.173417f, 94.47046f, -8.999331f, 108.6454f, -9.912288f, 115.7329f, -6.260461f, 87.38299f, -7.173417f, 94.47046f, -8.999331f, 108.6454f, -9.912288f, 115.7329f, -6.260461f, 87.38299f, -7.173417f, 94.47046f, -8.999331f, 108.6454f, -9.912288f, 115.7329f, -6.260461f, 87.38299f, -7.173417f, 94.47046f, -8.999331f, 108.6454f, -9.912288f, 115.7329f, -6.260461f, 87.38299f, -7.173417f, 94.47046f, -8.999331f, 108.6454f, -9.912288f, 115.7329f });
var actual = Matrix <float> .Build.Dense(batchSize, k *filterGridWidth *filterGridHeight);
ConvUtils.ReshapeConvolutionsToRowMajor(convolutedInput, inputDepth, inputHeight, inputWidth, filterHeight, filterWidth,
padding, padding, stride, stride, BorderMode.Valid, actual);
var expected = Matrix <float> .Build.Dense(batchSize, k *filterGridWidth *filterGridHeight, new float[] { -6.260461f, -6.260461f, -6.260461f, -6.260461f, -6.260461f, -7.173417f, -7.173417f, -7.173417f, -7.173417f, -7.173417f, -8.999331f, -8.999331f, -8.999331f, -8.999331f, -8.999331f, -9.912288f, -9.912288f, -9.912288f, -9.912288f, -9.912288f, 87.38299f, 87.38299f, 87.38299f, 87.38299f, 87.38299f, 94.47046f, 94.47046f, 94.47046f, 94.47046f, 94.47046f, 108.6454f, 108.6454f, 108.6454f, 108.6454f, 108.6454f, 115.7329f, 115.7329f, 115.7329f, 115.7329f, 115.7329f });
MatrixAsserts.AreEqual(expected, actual);
}
public void BatchNormalizationLayer_Forward_SpatialInput()
{
var batchSize = 2;
var filterHeight = 2;
var filterWidth = 2;
var filterDepth = 2;
var stride = 1;
var padding = 0;
var inputWidth = 3;
var inputHeight = 3;
var inputDepth = 3;
var filterGridWidth = ConvUtils.GetFilterGridLength(inputWidth, filterWidth,
stride, padding, BorderMode.Valid);
var filterGridHeight = ConvUtils.GetFilterGridLength(inputHeight, filterHeight,
stride, padding, BorderMode.Valid);
var k = filterDepth;
var input = new float[] { 111, 121, 112, 122, 113, 123, 114, 124, 211, 221, 212, 222, 213, 223, 214, 224 };
var convInput = Matrix <float> .Build.Dense(2, 8, input);
var rowWiseInput = Matrix <float> .Build.Dense(batchSize, k *filterGridWidth *filterGridHeight);
Trace.WriteLine(convInput);
ConvUtils.ReshapeConvolutionsToRowMajor(convInput, inputDepth, inputHeight, inputWidth,
filterHeight, filterWidth, padding, padding, stride, stride,
BorderMode.Valid, rowWiseInput);
Trace.WriteLine(rowWiseInput);
var sut = new BatchNormalizationLayer();
sut.Initialize(filterGridWidth, filterGridHeight, filterDepth, batchSize,
Initialization.GlorotUniform, new Random(232));
var actual = sut.Forward(rowWiseInput);
Trace.WriteLine(string.Join(", ", actual.ToColumnMajorArray()));
Trace.WriteLine(actual);
var expected = Matrix <float> .Build.Dense(batchSize, k *filterGridWidth *filterGridHeight, new float[] { -1.0297426f, 0.9697576f, -1.00974762f, 0.9897526f, -0.9897526f, 1.00974762f, -0.9697576f, 1.0297426f, -1.0297426f, 0.9697576f, -1.00974762f, 0.9897526f, -0.9897526f, 1.00974762f, -0.9697576f, 1.0297426f });
MatrixAsserts.AreEqual(expected, actual);
}
/// <summary>
///
/// </summary>
/// <param name="input"></param>
/// <returns></returns>
public Matrix <float> Forward(Matrix <float> input)
{
m_inputActivations = input;
// Arrange input item for GEMM version of convolution.
ConvUtils.Batch_Im2Col(m_inputActivations, InputDepth, InputHeight, InputWidth,
FilterWidth, FilterHeight, m_padHeight, m_padWidth, m_stride, m_stride, BorderMode, Im2Cols);
// matrix multiplication for convolution
Weights.Multiply(Im2Cols, Conv);
Conv.AddColumnWise(Bias, Conv);
// Return the covolved data to row major and copy data to output
ConvUtils.ReshapeConvolutionsToRowMajor(Conv, InputDepth, InputHeight, InputWidth,
FilterWidth, FilterHeight, m_padHeight, m_padWidth, m_stride, m_stride, BorderMode, OutputActivations);
return(OutputActivations);
}
/// <summary>
///
/// </summary>
/// <param name="delta"></param>
/// <returns></returns>
public Matrix <float> Backward(Matrix <float> delta)
{
// Reshape delta to fit with data layout in im2col
ConvUtils.ReshapeConvolutionsToRowMajor(delta, InputDepth, InputHeight, InputWidth,
FilterWidth, FilterHeight, m_padHeight, m_padWidth, m_stride, m_stride, BorderMode, m_deltaInReshape);
// Calculate gradients for weights and biases
m_deltaInReshape.TransposeAndMultiply(Im2Cols, WeightsGradients);
m_deltaInReshape.SumRows(BiasGradients);
// calcualte delta for next layer.
Weights.TransposeThisAndMultiply(m_deltaInReshape, Im2Cols);
// convert back to original layout
m_delta.Clear();
ConvUtils.Batch_Col2Im(Im2Cols, InputDepth, InputHeight, InputWidth,
FilterHeight, FilterWidth, m_padHeight, m_padWidth, m_stride, m_stride, BorderMode, m_delta);
return(m_delta);
}