public void backward()
{
int inSize, outSize;
var countExamples = 3;
inSize = outSize = 5;
var input = MatrixD.Build.Random(countExamples, inSize);
var layer = new BiasLayer(inSize);
var nextGradients = MatrixD.Build.Random(countExamples, outSize);
Assert.IsTrue(layer.backward(input, nextGradients).EEquals(nextGradients));
}
public void forward()
{
var countExamples = 3;
var inSize = 5;
var input = MatrixD.Build.Random(countExamples, inSize);
var layer = new BiasLayer(inSize);
var expectedOutputs = MatrixD.Build.repeat(countExamples, inSize, 0);
for (var i = 0; i < countExamples; i++)
{
for (var j = 0; j < inSize; j++)
{
expectedOutputs[i, j] = input[i, j] + layer.Biases[j];
}
}
Assert.IsTrue(layer.forward(input).EEquals(expectedOutputs));
}
public void backwardLearn()
{
var learnRate = GlobalRandom.NextDouble();
int inSize, outSize;
var countExamples = 3;
inSize = outSize = 5;
var input = MatrixD.Build.Random(countExamples, inSize);
var layer = new BiasLayer(inSize);
var nextGradients = MatrixD.Build.Random(countExamples, outSize);
var expectedBiases = layer.Biases.ShallowCopy();
var gradientSums = nextGradients.ColumnSums().AsArray();
for (int i = 0; i < expectedBiases.Length; i++)
{
expectedBiases[i] -= learnRate * gradientSums[i];
}
layer.backwardLearn(input, nextGradients, learnRate);
Assert.IsTrue(expectedBiases.EEquals(layer.Biases));
}