The EvaluateMVA() method of TMVA::Reader* class leaks memory when used with Neural Network model trained in Keras

ROOT Version: v6-28-02-5
Platform: Ubuntu 22.04.2 LTS
Compiler: (GCC) 12.2.0

Dear experts! I’ve been using ROOT for Particle Physics research and found one interensting pattern which don’t know how to resolve. I got ROOT as part of AliPhysics installation. I use PyKeras interface to train neural network via PyROOT using standart approach by prepairing the trees, using factory.BookMethod(dataloader, ROOT.TMVA.Types.kPyKeras, “PyKeras”, string_of_options) and then calling factory.TrainAllMethods(). From this I get a weights.xml file. The trained model is very basic:

model = keras.Sequential(
                keras.layers.Dense(128, activation="relu"), 
                keras.layers.Dense(128, activation="relu"),
                keras.layers.Dense(128, activation="relu"),
                keras.layers.Dense(2, activation="softmax"),

Since the code of my main task is written in C++ then I try to call trained network in it to obtain its inference results. The excerpt of the code where this is done is posted below. All the other code in my task does not interact with this excerpt in any way. I’ve changed fMVAvariable values for dummy values for simplicity, but the behaviour of the code does not change with this. fNNCandidateVariablesNames values are also unimportant in this context.
When I run the code with this excerpt the RAM usage steadily increases and after a couple of minutes takes all available RAM. If I train and then use for interence another methods, like BDT or neural networsk implementation TMVA::kDL the RAM usage does not increase. Also, if I comment out one line with EvaluateMVA(“KerasNN”); the memory usage is also okay, no leakage occurs. Important thing to note that when using the line with EvaluateMVA(“KerasNN”); the inference results are produced(I get the number which is networks prediction), so in this sense the code works, though with ever increasing RAM usage. I will really appreciate any comment on my situation.

  Int_t NumberOfVariables = 20;

  TString *fNNCandidateVariablesNames = new TString[NumberOfVariables];

  fNNCandidateVariablesNames[0] = "massCand";
  fNNCandidateVariablesNames[1] = "ptLb";
  fNNCandidateVariablesNames[2] = "pt_Prong0";
  fNNCandidateVariablesNames[3] = "pt_Prong1";
  fNNCandidateVariablesNames[4] = "d0_Prong0";
  fNNCandidateVariablesNames[5] = "cosThetaStar";
  fNNCandidateVariablesNames[6] = "Ct";
  fNNCandidateVariablesNames[7] = "Prodd0";
  fNNCandidateVariablesNames[8] = "cosp";
  fNNCandidateVariablesNames[9] = "cospXY";
  fNNCandidateVariablesNames[10] = "NormDL";
  fNNCandidateVariablesNames[11] = "ImpPar";
  fNNCandidateVariablesNames[12] = "dca";
  fNNCandidateVariablesNames[13] = "ptLc";
  fNNCandidateVariablesNames[14] = "d0_Prong0Lc";
  fNNCandidateVariablesNames[15] = "d0_Prong1Lc";
  fNNCandidateVariablesNames[16] = "pt_Prong0Lc";
  fNNCandidateVariablesNames[17] = "pt_Prong1Lc";
  fNNCandidateVariablesNames[18] = "cosv0pointinganglexyLc";
  fNNCandidateVariablesNames[19] = "normalizedv0decaylengthxyLc";

  for (Int_t icand = 0; icand < 100000; icand++) {

    TMVA::Reader* MyReader = new TMVA::Reader();
    Float_t *fMVAvariable = new Float_t[NumberOfVariables];

    MyReader->AddVariable(fNNCandidateVariablesNames[0], &fMVAvariable[0]);
    MyReader->AddVariable(fNNCandidateVariablesNames[1], &fMVAvariable[1]);
    MyReader->AddVariable(fNNCandidateVariablesNames[2], &fMVAvariable[2]);
    MyReader->AddVariable(fNNCandidateVariablesNames[3], &fMVAvariable[3]);
    MyReader->AddVariable(fNNCandidateVariablesNames[4], &fMVAvariable[4]);
    MyReader->AddVariable(fNNCandidateVariablesNames[5], &fMVAvariable[5]);
    MyReader->AddVariable(fNNCandidateVariablesNames[6], &fMVAvariable[6]);
    MyReader->AddVariable(fNNCandidateVariablesNames[7], &fMVAvariable[7]);
    MyReader->AddVariable(fNNCandidateVariablesNames[8], &fMVAvariable[8]);
    MyReader->AddVariable(fNNCandidateVariablesNames[9], &fMVAvariable[9]);
    MyReader->AddVariable(fNNCandidateVariablesNames[10], &fMVAvariable[10]);
    MyReader->AddVariable(fNNCandidateVariablesNames[11], &fMVAvariable[11]);
    MyReader->AddVariable(fNNCandidateVariablesNames[12], &fMVAvariable[12]);
    MyReader->AddVariable(fNNCandidateVariablesNames[13], &fMVAvariable[13]);
    MyReader->AddVariable(fNNCandidateVariablesNames[14], &fMVAvariable[14]);
    MyReader->AddVariable(fNNCandidateVariablesNames[15], &fMVAvariable[15]);
    MyReader->AddVariable(fNNCandidateVariablesNames[16], &fMVAvariable[16]);
    MyReader->AddVariable(fNNCandidateVariablesNames[17], &fMVAvariable[17]);
    MyReader->AddVariable(fNNCandidateVariablesNames[18], &fMVAvariable[18]);
    MyReader->AddVariable(fNNCandidateVariablesNames[19], &fMVAvariable[19]);

    MyReader->BookMVA("KerasNN", TString("./dataset/weights/Keras_model_PyKeras.weights.xml"));

    /* MyReader->BookMVA("BDT", TString("./dataset/weights/Keras_model_BDT.weights.xml")); */

    fMVAvariable[0] = 1;
    fMVAvariable[1] = 1;
    fMVAvariable[2] = 1;
    fMVAvariable[3] = 1;
    fMVAvariable[4] = 1;
    fMVAvariable[5] = 1;
    fMVAvariable[6] = 1;
    fMVAvariable[7] = 1;
    fMVAvariable[8] = 1;
    fMVAvariable[9] = 1;
    fMVAvariable[10] = 1;
    fMVAvariable[11] = 1;
    fMVAvariable[12] = 1;
    fMVAvariable[13] = 1;
    fMVAvariable[14] = 1;
    fMVAvariable[15] = 1;
    fMVAvariable[16] = 1;
    fMVAvariable[17] = 1;
    fMVAvariable[18] = 1;
    fMVAvariable[19] = 1;

    Double_t fMVAResponse_Sig = MyReader->EvaluateMVA("KerasNN");

    /* Double_t fMVAResponse_Sig = MyReader->EvaluateMVA("BDT"); */

    delete MyReader;
    delete[] fMVAvariable; 
    fMVAvariable = nullptr;


Welcome to the ROOT Forum!
I’m sure @moneta will be able to help

Dear @moneta, if you have time I’ll really appreciate if you take a look at this sutuation.

Sorry for my late reply. It is not clear 100% to me how you are using the class. Are you creating a Reader object for every single evaluation ? In principle you should create the Reader ones and then loop on your entries to perform the evaluation. This case is not leaking the memory. Instead creating the Reader every time it could be different, since this requires to instantiate the Keras TMVA Method class, which being instantiates some Python interpreter objects. This is possible is leaking, I would need to further investigate, but it is not designed to be called for every event, given the large overhead in creating that class

Best regards


@moneta, thank you for the answer. As I wrote above, I have a big file with C++ code and the code I posted here is just a small part of it. Because of how the other code in a big file is written I have no choise but to create, use and then delete a new Reader class instance for every event I process and there are hundreds of them. So in the code I pasted above I created an artificial situation where I create, use and delete Reader instances in a loop which is logically equivalent to the code written in a big file. To be sure that the problem persists, I also confirmed that the increasing RAM usage occurs if I paste the code from above in the very begining of the big file, where the execution just starts.
But what’s bothering me the most is that if I create Reader class instance in the begining of the loop and then delete it in the end of the loop without using EvaluateMVA() then no leakage occurs. I don’t know whether this could be explained by the reason you wrote above(just because of my lacking knowledge of ROOT), but if that is so it’ll be good to know.
Just in case, my tensorflow and keras versions are 2.12.0

I don’t understand why you need to create a Reader for every event. You have a different model for every event ?
I will investigate anyway why you have a big leak by creating a Reader for every event