public void LossDeriv(Matrix output, Matrix expectedOutput, Matrix result, float regularizationParam)
{
#if GPU
KernelManager.Loss(expectedOutput, output, result, func_deriv);
#elif CPU
switch (func_deriv)
{
case "quadratic_loss_deriv":
{
Parallel.For(0, result.Rows, (i) => result.Memory[i] += (output.Memory[i] - expectedOutput.Memory[i]) / result.Rows + regularizationParam);
}
break;
case "binary_cross_entropy_deriv":
{
//- (z / (y + eps) - (1 - z) / (1 - y + eps));
//z = expected out
//y = actual output
Parallel.For(0, result.Rows, (i) =>
{
float r_0 = ((1.0f - expectedOutput.Memory[i]) - output.Memory[i]);
float r = (r_0 == 0) ? 0 : (1.0f / r_0);
result.Memory[i] += r + regularizationParam;
//result.memory[i] += -(float)((expectedOutput.memory[i] / (output.memory[i] + float.Epsilon) - (1 - expectedOutput.memory[i]) / (1 - output.memory[i] + float.Epsilon))/result.Rows);
});
}
break;
case "gan_disc_fake":
{
//- (z / (y + eps) - (1 - z) / (1 - y + eps));
//z = expected out
//y = actual output
Parallel.For(0, result.Rows, (i) =>
{
float r_0 = (1.0f - expectedOutput.Memory[i] - output.Memory[i]);
float r = (r_0 == 0) ? 0 : (1.0f / r_0);
result.Memory[i] += r + regularizationParam;
//result.memory[i] += -(float)((expectedOutput.memory[i] / (output.memory[i] + float.Epsilon) - (1 - expectedOutput.memory[i]) / (1 - output.memory[i] + float.Epsilon))/result.Rows);
});
}
break;
case "gan_disc_real":
{
//- (z / (y + eps) - (1 - z) / (1 - y + eps));
//z = expected out
//y = actual output
Parallel.For(0, result.Rows, (i) =>
{
float r_0 = ((1 - expectedOutput.Memory[i]) + output.Memory[i]);
float r = (r_0 == 0) ? 0 : -(1.0f / r_0);
result.Memory[i] += r + regularizationParam;
//result.memory[i] += -(float)((expectedOutput.memory[i] / (output.memory[i] + float.Epsilon) - (1 - expectedOutput.memory[i]) / (1 - output.memory[i] + float.Epsilon))/result.Rows);
});
}
break;
case "gan_gen":
{
//- (z / (y + eps) - (1 - z) / (1 - y + eps));
//z = expected out
//y = actual output
Parallel.For(0, result.Rows, (i) =>
{
float r_0 = -1.0f / (output.Memory[i]);
result.Memory[i] += r_0 + regularizationParam;
//result.memory[i] += -(float)((expectedOutput.memory[i] / (output.memory[i] + float.Epsilon) - (1 - expectedOutput.memory[i]) / (1 - output.memory[i] + float.Epsilon))/result.Rows);
});
}
break;
}
#endif
}
public void Loss(Matrix output, Matrix expectedOutput, Matrix result, float regularizationParam)
{
#if GPU
KernelManager.Loss(expectedOutput, output, result, func);
#elif CPU
switch (func)
{
case "quadratic_loss":
{
Parallel.For(0, result.Rows, (i) => result.Memory[i] += (0.5f * (float)Math.Pow(output.Memory[i] - expectedOutput.Memory[i], 2)) / result.Rows + regularizationParam);
}
break;
case "binary_cross_entropy":
{
//- (z * log(y + eps) + (1-z) * log(1 - y + eps))
//z = expected out
//y = actual output
Parallel.For(0, result.Rows, (i) =>
{
result.Memory[i] += (-(float)(expectedOutput.Memory[i] * Math.Log(output.Memory[i] + float.Epsilon) + (1 - expectedOutput.Memory[i]) * Math.Log(1 - output.Memory[i] + float.Epsilon)) / result.Rows + regularizationParam);
});
}
break;
case "gan_disc_fake":
{
//- (z / (y + eps) - (1 - z) / (1 - y + eps));
//z = expected out
//y = actual output
Parallel.For(0, result.Rows, (i) =>
{
float r_0 = (float)Math.Log(1.0f - expectedOutput.Memory[i] - output.Memory[i]);
result.Memory[i] += r_0 + regularizationParam;
//result.memory[i] += -(float)((expectedOutput.memory[i] / (output.memory[i] + float.Epsilon) - (1 - expectedOutput.memory[i]) / (1 - output.memory[i] + float.Epsilon))/result.Rows);
});
}
break;
case "gan_disc_real":
{
//- (z / (y + eps) - (1 - z) / (1 - y + eps));
//z = expected out
//y = actual output
Parallel.For(0, result.Rows, (i) =>
{
float r_0 = -(float)Math.Log((1 - expectedOutput.Memory[i]) + output.Memory[i]);
result.Memory[i] += r_0 + regularizationParam;
//result.memory[i] += -(float)((expectedOutput.memory[i] / (output.memory[i] + float.Epsilon) - (1 - expectedOutput.memory[i]) / (1 - output.memory[i] + float.Epsilon))/result.Rows);
});
}
break;
case "gan_gen":
{
//- (z / (y + eps) - (1 - z) / (1 - y + eps));
//z = expected out
//y = actual output
Parallel.For(0, result.Rows, (i) =>
{
float r_0 = -(float)Math.Log(output.Memory[i]);
result.Memory[i] += r_0 + regularizationParam;
//result.memory[i] += -(float)((expectedOutput.memory[i] / (output.memory[i] + float.Epsilon) - (1 - expectedOutput.memory[i]) / (1 - output.memory[i] + float.Epsilon))/result.Rows);
});
}
break;
}
#endif
}