cureos / csmpfit Goto Github PK

View Code? Open in Web Editor NEW

30.0 30.0 19.0 126 KB

A Least Squares Optimization Library in C#

C# 100.00%

csmpfit's People

Contributors

Stargazers

Watchers

Forkers

gitter-badger msevestre dcuccia frankliu-x kouweizhong lulzzz svengl louiszhuang did4730 beatskip stec0747 lazicsasalaki fabiannitsche rubyebrown domenicohyatt moosehat-mux oniesmitham natwijaya19

csmpfit's Issues

Failure when first parameter has analytic derivative and next does not.

It seems there may be a bookkeeping issue indexing the parameter derivatives when a parameter with user-supplied analytic derivatives is followed by one with numerical derivatives. This issue was observed after implementing the fix from the previous issue (#3), so there's a chance it could be due to that change. It seems the following change to line 1294 of MPFit.cs fixes the problem:

if (dside != null && dsidei == 3)
{
  ij += m; // bug fix
  continue;
}
//if (dside != null && dsidei == 3) continue; // replaced

The issue can be demonstrated by adding the following test to TestMPFit.cs:

   private static void TestGaussianWithDerivs()
   {
        double[] x = {-1.7237128E+00,1.8712276E+00,-9.6608055E-01,
    -2.8394297E-01,1.3416969E+00,1.3757038E+00,
    -1.3703436E+00,4.2581975E-02,-1.4970151E-01,
    8.2065094E-01};
        double[] y = {-4.4494256E-02,8.7324673E-01,7.4443483E-01,
    4.7631559E+00,1.7187297E-01,1.1639182E-01,
    1.5646480E+00,5.2322268E+00,4.2543168E+00,
    6.2792623E-01};
        double[] ey = new double[10];
        double[] p = { 0.0, 1.0, 1.0, 1.0 };       /\* Initial conditions _/
        double[] pactual = { 0.0, 4.70, 0.0, 0.5 };/_ Actual values used to make data_/
                                                   //double[] perror = new double[4];              /_ Returned parameter errors _/
        mp_par[] pars = new mp_par[4] /_ Parameter constraints */
                            {
                                new mp_par() { side = 3, deriv_debug = 0 },
                                new mp_par() { side = 1, deriv_debug = 0 },
                                new mp_par() { side = 3, deriv_debug = 0 },
                                new mp_par() { side = 1, deriv_debug = 0 }
                                //new mp_par() { deriv_debug = 1 },
                                //new mp_par() { deriv_debug = 1 },
                                //new mp_par() { deriv_debug = 1 },
                                //new mp_par() { deriv_debug = 1 }
                            };

    mp_result result = new mp_result(4);
    //result.xerror = perror;

    /* No constraints */

    for (uint i = 0; i < 10; i++) ey[i] = 0.5;

    CustomUserVariable v = new CustomUserVariable() { X = x, Y = y, Ey = ey };

    /* Call fitting function for 10 data points and 4 parameters (no
       parameters fixed) */

    var logger = new System.IO.StringWriter();
    int status = MPFit.Solve(GaussianFuncAndDerivs, 10, 4, p, pars, null, v, ref result, logger);

    Console.Write("*** TestGaussFitWithDerivs status = {0}\n", status);
    PrintResult(p, pactual, result);
    Console.WriteLine(logger.ToString());
}

private static int GaussianFuncAndDerivs(double[] p, double[] dy, IList<double>[] dvec, object vars)
{
    CustomUserVariable v = (CustomUserVariable)vars;
    double[] x, y, ey;
    double sig2;

    x = v.X;
    y = v.Y;
    ey = v.Ey;

    sig2 = p[3] * p[3];
    //var dvec1 = new double[dy.Length];
    //var residuals = new double[dy.Length];

    for (int i = 0; i < dy.Length; i++)
    {
        double xc = x[i] - p[2];
        double exp = Math.Exp(-0.5 * xc * xc / sig2);
        //residuals[i] = (y[i] - p[1] * exp - p[0]) / ey[i];
        dy[i] = (y[i] - p[1] * exp - p[0]) / ey[i];

        // NOTE: it would make sense to store the first 2 derivatives in vars since they don't change.
        if (dvec != null)
        {
            if (dvec[0] != null)
            {
                dvec[0][i] = -1.0 / ey[i];
            }
            if (dvec[1] != null)
            {
                //dvec1[i] = -exp / ey[i];
                dvec[1][i] = -exp / ey[i];
            }
            if (dvec[2] != null)
            {
                dvec[2][i] = -p[1] * xc * exp / (ey[i] * sig2);
            }
            if (dvec[3] != null)
            {
                dvec[3][i] = -p[1] * xc * xc * exp / (ey[i] * p[3] * sig2);
            }
        }
    }

    // Array assignment rather than element-wise causes failure.
    //dy = residuals;

    // Array mismatch exception
    //if (dvec != null)
    //{
    //  if (dvec[1] != null)
    //  {
    //      dvec[1] = dvec1;
    //  }
    //}
    return 0;
}

multiple curves fitting

I have two curves described by one model, and I want to fit with them simultaneously. How to fit for multiple curves?

maxiter==0 is not working as expected (e.g. should turn off fitting)

The documentation for maxiter parameter specifies

/* Maximum number of iterations. If maxiter == 0,
then basic error checking is done, and parameter
errors/covariances are estimated based on input
parameter values, but no fitting iterations are done. */

However it seems that setting the value to 0 is ignored by the code when the user values are being read

see

csmpfit/src/MPFitLib/MPFit.cs

Line 378 in 5355183

if (config.maxiter > 0) conf.maxiter = config.maxiter;

I think it should read
if (config.maxiter >= 0) conf.maxiter = config.maxiter;

instead of
if (config.maxiter > 0) conf.maxiter = config.maxiter;

Nullable reference support

Array mismatch exception when using analytical derivatives

It would be good if something testing user-supplied analytical derivatives were included in the unit tests. I expected to see them in ForwardModels.GaussFunc().

The following change to line 1219 in MPFit.cs seems to prevent the exception on line 1242 (dvec[ifree[j]] = ....).

dvec = new DelimitedArray<double>[npar];    // Prevents array type mismatch error on dvec[ifree[j]] = new DelimitedArray<double>(fjac, j * m, m) -RDN 6/17/2016
//dvec = new double[npar][];

Adding the following test to TestMPFit.cs illustrates the issue:

private static void TestAnalyticDerivatives()
{
        const double intercept = 7.0;
        const double slope = -3.0;
        var pactual = new double[] { intercept, slope };
        var p = new double[] { 5.9, -1.1 };
        var x = new double[] { -5, -3, -1, 0, 1, 3, 5 };
        var y = new double[x.Length];


    for (uint i = 0; i < x.Length; i++)
    {
        y[i] = intercept + slope * x[i];
    }

    var mpPar = new mp_par[] { new mp_par(), new mp_par() };
    mpPar[0].side = 3;
    mpPar[1].side = 3;

    var result = new mp_result(2);

    var status = MPFit.Solve(LineFunc, x.Length, 2, p, mpPar, null, new LineFitData(x, y), ref result);

    Console.Write("*** TestLineFit status = {0}\n", status);
    PrintResult(p, pactual, result);
}

private class LineFitData
{
    public LineFitData(double[] x, double[] y)
    {
        this.X = x;
        this.Y = y;
    }

    public double[] X { get; }
    public double[] Y { get; }
}

private static int LineFunc(double[] p, double[] dy, IList<double>[] dvec, object vars)
{
    var lineFitData = (LineFitData)vars;
    var x = lineFitData.X;
    var y = lineFitData.Y;

    for (int i = 0; i < x.Length; i++)
    {
        dy[i] = y[i] - (p[0] + p[1] * x[i]);
        if (dvec != null)
        {
            var dvec0 = dvec[0];
            var dvec1 = dvec[1];
            if (dvec0 != null)
            {
                dvec0[i] = -p[0];
            }
            if (dvec1 != null)
            {
                dvec1[i] = -x[i];
            }
        }
    }
    return 0;
}

Element-wise assignment appears to be required in user-supplied function.

In the user-supplied function to be fitted, it seems that both the residuals (dy) and derivatives (dvec) need to be set element-by-element, as opposed to array assignment. dy = array seems to have no effect on output, and dvec[i] = array results in an array-type mismatch exception. Commented out code in the test included in issue #4 demonstrates the problem. Also (possibly unrelated), would it make sense to use the ref keyword in the user function?

Documentation

Could you please provide a documentation for this library(C#)?

Strong Name is not valid from NUGET

I am unable to publish and install a project using ClickOnce due to mpfitlib.dll failing the strong name check, This is with the latest NUGET package v1.1.0

To confirm this, I ran sn.exe -vf mpfitlib.dll which also return that it was invalid.

I then downloaded the latest source from Github and built a release version. The resulting dll passed the strong name verification. This suggests to me that something is wrong with version currently uploaded to NUGET.

Would you be interested in a PR to add huber loss function?

I am planning on modifying the library to support a huber loss function. Would you be interested in the eventual pull request? If not, I won't bother with all of the github stuff.

Replace PCL profile 328 with .NET Standard 2.0

mp_result.covar and mp_result.resid seems to be unassigned

It may be because I do not know how to use it well, but 'mp_result.covar' and 'mp_result.resid' seem to be unassigned.. how to access these parameters?

Provide .NET Standard library

Recommend we add a .NET Standard library, which allows us to support multiple platforms with a single binary.

https://docs.microsoft.com/en-us/dotnet/standard/library

Per the above, we should provide a NETStandard1.0 library to maximize the runtimes this will work with, including .NET Framework 4.5+, .NET Core 1.0+, Mono 4.6+, Xamarin.iOS 10+, Xamarin.Android 7+, and UWP 10+.

I've created a PR here for review: #10