unvalid xml file and other basic questions about SVM trainers

[CARL-Bouvet]

Hello guys, hello Vincent,
I ask noob questions, sorry about that.

My xml file for a very simple detection set return that common error :

terminate called after throwing an instance of 'std::runtime_error'
  what():  OpenCV Error: (The node does not represent a user object (unknown type?)) in cvRead, in file /home/.../node_modules/opencv-build/opencv/ope
ncv/modules/core/src/persistence_c.cpp, line 1426, status -2
Abandon (core dumped)

It is probably the SVM trainer that gets me an invalid XML file.
So, I have a lot of questions before I can to be on my own:

At first; how to get settings chosen by the computer if we use the trainauto mode ?
Is there a command for console.log theme? even in python, this command is hard to find.

The code of my SVM trainer is almost the same as @justadudewhohacks 's one.

Thank you all.

[justadudewhohacks]

Hm not sure if the actual error is related to this package, rather then being an issue with OpenCV not being able to load your xml file.

To your second question, when using trainAuto, shouldn't you be able to get all the updated properties by logging the svm object: console.log(svm)?

[CARL-Bouvet]

thank you for your answer.

console.log(svm)

it's as simple as that!? shame on me...
ok, that command works:

SVM {
  isTrained: true,
  varCount: 10404,

  classWeights:
   Mat {
     step: 0,
     elemSize: 0,
     sizes: [],
     empty: true,
     depth: 0,
     dims: 0,
     channels: 1,
     type: 0,
     cols: 0,
     rows: 0 },
  kernelType: 2,
  p: 0,
  nu: 0,
  gamma: 0.00015000000000000001,
  degree: 0,
  coef0: 0,
  c: 312.5 }

But average prediction results with test datas is 0.8208955223880597 so WTF*$#?!
My trainData mats are not empty so my problême is probably my HOG paramêtres, or letters in train pictures are too small or I don t know ...

const hog = new cv.HOGDescriptor({
  winSize: new cv.Size(90, 90),
  blockSize: new cv.Size(10, 10),
  blockStride: new cv.Size(5, 5),
  cellSize: new cv.Size(5, 5),
  L2HysThreshold: 0.2,
  nbins: 9, 
  gammaCorrection: true,
  signedGradient: true 
});

If you see something wrong, please tell me. I continue my investigation.
I will probably ask other questions soon so don't close that tread please. and thank you again.

[justadudewhohacks]

Uhm 0.82 is pretty good, that's overall about what I achieved in the example, isn't it?

[CARL-Bouvet]

yes. but with my own data.

const arrayDAlphabetMinus = Array(26).fill(97).map((v, i) => v + i).map(ascii => String.fromCharCode(ascii));
const arrayDAlphabetMajus = Array(26).fill(65).map((v, i) => v + i).map(ascii => String.fromCharCode(ascii));
const arrayNombre = Array(10).fill(48).map((v, i) => v + i).map(ascii => String.fromCharCode(ascii));
const arrayPonct = ([33, 39, 44, 45, 63]).map(ascii => String.fromCharCode(ascii));
const arrayDeToutLesCaracteres = ((arrayDAlphabetMinus.concat(arrayDAlphabetMajus)).concat(arrayNombre)).concat(arrayPonct);

const data = arrayDeToutLesCaracteres.map((char) => {
  const trainDataDir = `${trainDataPath}/${char}`;
  const testDataDir = `${testDataPath}/${char}`;
  const train = fs.readdirSync(trainDataDir).map(file => `${trainDataDir}/${file}`);
  const test = fs.readdirSync(testDataDir).map(file => `${testDataDir}/${file}`);
  return ({ train, test });
});

I think i find my mistake: our train pictures must be on THRESH_BINARY (or THRESH_BINARY_INV) for to be readable by the HOG?

[CARL-Bouvet]

no, it still does not work. I'm stalling!

(function main(){
  const img = './testpourSVMtraining.png';
  const xmlFilePath = ('./SVMdetest.xml');
  const classifier = new cv.CascadeClassifier(xmlFilePath); /* it doesn t want to use this xml file. */

  const src = cv.imread(img);
  const letterRects = classifier.detectMultiScale(src).objects;
  if (!letterRects.length){
    throw new Error("nothing detected.");
  }
  const result1 = img.getRegion(letterRects[0]);
  console.log(result1);
 })();

terminate called after throwing an instance of 'std::runtime_error'
  what():  OpenCV Error: (The node does not represent a user object (unknown type?)) in cvRead, in file /home/.../node_modules/opencv-build/opencv/opencv/modu
les/core/src/persistence_c.cpp, line 1426, status -2
Abandon (core dumped)

some train images:

and My HOGAndSVMTrainer.js

const fs = require('fs');
const cv = require('./node_modules/opencv4nodejs');
const path = require('path')
const trainDataPath = './trainData';
const testDataPath = './testData';
const outPath = './';
const SVMFile = 'SVMdetest.xml';

const hog = new cv.HOGDescriptor({
  winSize: new cv.Size(90, 90),
  blockSize: new cv.Size(60, 60),
  blockStride: new cv.Size(30, 30),
  cellSize: new cv.Size(30, 30),
  L2HysThreshold: 0.2,
  nbins: 9, 
  gammaCorrection: true,
  signedGradient: true 
});
const svm = new cv.SVM({
  kernelType: cv.ml.SVM.RBF,
  c: 12.5,
  gamma: 0.50625 
});

const arrayDAlphabetMinus = Array(26).fill(97).map((v, i) => v + i).map(ascii => String.fromCharCode(ascii));
const arrayDAlphabetMajus = Array(26).fill(65).map((v, i) => v + i).map(ascii => String.fromCharCode(ascii));
const arrayNombre = Array(10).fill(48).map((v, i) => v + i).map(ascii => String.fromCharCode(ascii));
const arrayPonct = ([33, 39, 44, 45, 63]).map(ascii => String.fromCharCode(ascii));
const arrayDeToutLesCaracteres = ((arrayDAlphabetMinus.concat(arrayDAlphabetMajus)).concat(arrayNombre)).concat(arrayPonct);
const data = arrayDeToutLesCaracteres.map((char) => {
  const trainDataDir = `${trainDataPath}/${char}`;
  const testDataDir = `${testDataPath}/${char}`;
  const train = fs.readdirSync(trainDataDir).map(file => `${trainDataDir}/${file}`);
  const test = fs.readdirSync(testDataDir).map(file => `${testDataDir}/${file}`);
  return ({ train, test });
});

const computeHOGDescriptorFromImage = (img) => {
 if (img.rows !== 90 || img.cols !== 90) {
    console.log(`l image ${img} n est pas au bon format ?! \n elle fait ${img.rows}x${img.cols}`);
    return 0;
  }
  return hog.compute(img);
};

const trainSVM = (trainDataFiles) => {
  console.log('constitution des données du HOG et étiquetage');
  const échantillonDImages = [];
  const étiquettes = [];
  trainDataFiles.forEach((files, label) => {
    files.forEach((file) => {
      const img = cv.imread(file);
      const desc = computeHOGDescriptorFromImage(img);
            if (!desc) {
        return;
      }

      échantillonDImages.push(desc);
      étiquettes.push(label);
    });
  });

  console.log('entrainement');
  const trainData = new cv.TrainData(
    new cv.Mat(échantillonDImages, cv.CV_32F),
    0,
    new cv.Mat([étiquettes], cv.CV_32S)
  );
  svm['trainAuto'](trainData);
  console.log(svm);
};

const trainDataFiles = data.map(classData => classData.train); /*data l35 */
const testDataFiles = data.map(classData => classData.test);

const numTrainImagesPerClass = trainDataFiles[0].length;
const numTestImagesPerClass = testDataFiles[0].length;
console.log('train data per class:', numTrainImagesPerClass);
console.log('test data per class:', numTestImagesPerClass);

trainSVM(trainDataFiles, true);
svm.save(`${outPath}/${SVMFile}`);
svm.load(`${outPath}/${SVMFile}`);

... without errs module, saveConfusionMatrix etc...
because I m not familiar with xml file, I look for in that direction now but if you see something, have a suggestion, I m all ears.

[justadudewhohacks]

Odd, could you upload the xml file here, which is being produced from svm.save?

[CARL-Bouvet]

sure!
upload link

or

SVMdetest.zip

[CARL-Bouvet]

I just tryed with a copy/paste of haarcascade_eye.xml and it's work perfectly.
So my problem is my code or my config.