Machine Learning toolbox and C Code generation: where do the LDA coefficients come from?

Question

handyman 2023-10-5

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/2029649-machine-learning-toolbox-and-c-code-generation-where-do-the-lda-coefficients-come-from

评论： handyman 2023-10-6

I have exported an LDA classifier to C code (5 classes, 8 features, diagonal) and as far as I can tell there are four vectors/matrices in the generated C code that contain the classifier coefficients:

in the "myclassifiername_predict" function:

C code:
void myclassifiername_predict(const double X[8], cell_wrap_0 label[1],double score[5])
{
    static const double dv[8] = {22.242777113313874, 10091.972043688236,
    18.120065167575063, 1.6703515441084688E+6,
    24.467860977243351, 10578.886659888578,
    19.515843697095619, 1.9037831242647208E+6};
    
    static const double dv1[5] = {0.19444444444444445, 0.19791666666666666,
    0.20254629629629631, 0.18865740740740741,
    0.21643518518518517};
    
    // rest of function body...
}
void DiagonalDiscriminant(const double X[8], double mah[5])
{
  static const double dv[40] = {
      92.058756510416757,    428.55436769005888,    99.445223214285761,
      121.16319976993877,    28.000396891711258,    12736.000000000013,
      81355.60233918138,     19612.525714285723,    16993.374233128856,
      3436.1497326203244,    143.11044456845252,    72.791255482456208,
      144.52926339285719,    196.37662960122714,    29.916151403743349,
      5.2292266666666726E+6, 3.3097596257309979E+6, 7.8175085714285746E+6,
      7.0232076564417249E+6, 894724.10695187247,    480.39411272321479,
      26.045367324561425,    53.19522321428574,     239.98229006901863,
      20.971883355614992,    80371.8095238096,      2976.1871345029267,
      8478.720000000003,     52414.036809816,       2602.4385026737991,
      325.64160156250028,    34.383543494152079,    84.338125000000034,
      72.35264091257676,     33.127548462566885,    1.6916089904761918E+7,
      1.5491024093567266E+6, 3.7583959771428583E+6, 4.3856852024539923E+6,
      1.6065081497326223E+6};
  static const double dv1[8] = {0.044958414810596169, 9.9088661331104676E-5,
                                0.0551874394905295,   5.9867637056829179E-7,
                                0.040869939588510124, 9.4527905643572937E-5,
                                0.051240418580971815, 5.2526991507303129E-7};
                                
   // rest of function body...                                  
}

When I look at the LDA classifier object in Matlab, I can see that

dv (in the predict function) is "d" in the compact Matlab classifier object
dv1 (in the predict function) is "classWeights" in the compact Matlab classifier object ("Prior" in the non-compact model)
dv (in the "DiagonalDiscrimant" function) is "mu" in the compact Matlab classifier object (5x8 Matrix)

What I can't figure out is where dv1 (in the "DiagonalDiscrimant" function) comes from?

The compact classifier object still has the 1x8 vector s and 8x8 matrix v, dv1 might be derived from that?

compactStruct.Impl.s =

1.4238

1.1705

1.0736

1.0486

0.9416

0.7696

0.6899

0.6289

compactStruct.Impl.v =