public void ReLUTests()
{
var input = new TensorOld(new double[] { 1, 2, -3, 5, -2, 7, 4, 6, 8, -5, 4, 1 }, 3, 4);
var expected = new TensorOld(new double[] { 1, 2, 0, 5, 0, 7, 4, 6, 8, 0, 4, 1 }, 3, 4);
var actual = TensorOld.Apply(input, Functions.ReLU);
Assert.Equal(expected, actual);
}
public void ReLUTests()
{
var output = new TensorOld(new double[] { 1, 2, 0, 5, 0, 7, 4, 6, 8, 0, 4, 1 }, 3, 4);
var expected = new TensorOld(new double[] { 1, 1, 0, 1, 0, 1, 1, 1, 1, 0, 1, 1 }, 3, 4);
var actual = TensorOld.Apply(output, Derivatives.ReLU);
Assert.Equal(expected, actual);
}
public void ApplyTest()
{
var t = TensorOld.Rand(20, 30);
var n = TensorOld.Apply(t, a => a * a);
for (int i = 0; i < 20; i++)
{
for (int j = 0; j < 30; j++)
{
Assert.Equal(t[i, j] * t[i, j], n[i, j]);
}
}
}
public void Optimize(TensorOld target, TensorOld gradient)
{
if (!dict.ContainsKey(gradient))
{
dict[gradient] = new AdamCache(gradient.Shape);
}
var c = dict[gradient];
TensorOld.Apply(c.M, gradient, c.M, (m, g) => Beta1 * m + (1 - Beta1) * g);
TensorOld.Apply(c.V, gradient, c.V, (v, g) => Beta2 * v + (1 - Beta2) * g * g);
TensorOld.Apply(c.M, c.V, c.T, (m, v) => Alpha * m / (Math.Sqrt(v) + E));
target.Minus(c.T);
}
public void Optimize(TensorOld target, TensorOld gradient)
{
if (!last.ContainsKey(gradient))
{
last[gradient] = gradient.GetSameShape();
TensorOld.Apply(gradient, last[gradient], g => LearningRate * g);
target.Minus(last[gradient]);
return;
}
var prev = last[gradient];
TensorOld.Apply(prev, gradient, prev, (p, g) => g * LearningRate - p * Moment);
target.Minus(prev);
}
public TensorOld Forward(TensorOld input)
{
input.Apply(a =>
{
if (a > Max)
{
return(Max);
}
if (a < Min)
{
return(Min);
}
return(a);
});
return(input);
}
public (TensorOld, TensorOld) GetXorData(int count)
{
var xbuff = DataEmulator.Instance.RandomArray(count, 2);
var x = new TensorOld(xbuff);
var y = new TensorOld(count, 1);
x.Apply(a => a * 12 - 6);
for (int i = 0; i < count; i++)
{
if (x[i, 0] * x[i, 1] > 0)
{
y[i, 0] = 0;
}
else
{
y[i, 0] = 1;
}
}
return(x, y);
}
public void Regularize(TensorOld parameters, TensorOld gradient)
{
TensorOld.Apply(gradient, parameters, gradient, (a, b) => a + Strength * b);
}
public void Normalize(TensorOld input, TensorOld output)
{
TensorOld.Apply(input, output, a => Math.Log10(a) / Denom);
}
/// <summary>
/// 反向传播或叫向前传播
/// </summary>
/// <param name="y">传回来的误差</param>
/// <returns>传到前面的误差</returns>
public override TensorOld Backward(TensorOld y)
{
TensorOld.Apply(ForwardOutput, y, BackwardOutput, (a, b) => a - b);
return(BackwardOutput);
}
/// <summary>
/// 反向传播或叫向前传播
/// </summary>
/// <param name="error">传回来的误差</param>
/// <returns>传到前面的误差</returns>
public override TensorOld Backward(TensorOld error)
{
TensorOld.Apply(ForwardOutput, Derivative, Derivatives.SigmoidFromOutput);
TensorOld.MultipleElementWise(Derivative, error, BackwardOutput);
return(BackwardOutput);
}
/// <summary>
/// 正向传播或叫向后传播
/// </summary>
/// <param name="input">输入的数值</param>
/// <returns>输出的数值</returns>
public override TensorOld Forward(TensorOld input)
{
TensorOld.Apply(input, ForwardOutput, Functions.Sigmoid);
return(ForwardOutput);
}
public void Normalize(TensorOld input, TensorOld output)
{
TensorOld.Apply(input, output, a => (a - Min) / Denom);
}
public void Normalize(TensorOld input, TensorOld output)
{
TensorOld.Apply(input, output, a => (a - Mean) / Delta);
}