內存泄漏可能是由遞歸函數造成的

我在C++中有一個隨機森林的實現，我通過mex在matlab中運行。它會順利運行，直到它達到下面的功能，它會卡住，並開始消耗內存，直到計算機凍結。內存泄漏可能是由遞歸函數造成的

void MyFunction(
    const IDataPointCollection& data, 
    std::vector<std::vector<int> >& leafNodeIndices, 
    ProgressStream* progress=0) const 
{ 
    ProgressStream defaultProgressStream(std::cout, Interest); 
    progress = (progress==0)?&defaultProgressStream:progress; 

    leafNodeIndices.resize(TreeCount()); 

    tbb::parallel_for<int>(0,TreeCount(),[&](int t) 
    { 
    leafNodeIndices[t].resize(data.Count()); 

    (*progress)[Interest] << "\rApplying tree " << t << "..."; 
    trees_[t]->Apply(data, leafNodeIndices[t]); 
    }); 
    (*progress)[Interest] << "STUCK HERE" << std::endl; 
    return; 
}

通過上面trees_[t]->Apply()的代碼去，我能夠將它縮小到下面的遞歸函數：

void ApplyNode(
    int nodeIndex, 
    const IDataPointCollection& data, 
    std::vector<unsigned int>& dataIndices, 
    int i0, 
    int i1, 
    std::vector<int>& leafNodeIndices, 
    std::vector<float>& responses_) 
{ 
std::cout<<"applying node"<<std::endl; 
    assert(nodes_[nodeIndex].IsNull()==false); 

    Node<F,S>& node = nodes_[nodeIndex]; 

    if (node.IsLeaf()) 
    { 
    for (int i = i0; i < i1; i++) 
     leafNodeIndices[dataIndices[i]] = nodeIndex; 
    return; 
    } 

    else if (i0 == i1) // No samples left 
    return; 

    else 
    { 
     for (int i = i0; i < i1; i++) 
      responses_[i] = node.Feature.GetResponse(data, dataIndices[i]); 

     int ii = Partition(responses_, dataIndices, i0, i1, node.Threshold); 

     // Recurse for child nodes. 
     ApplyNode(nodeIndex * 2 + 1, data, dataIndices, i0, ii, leafNodeIndices, responses_); 
     ApplyNode(nodeIndex * 2 + 2, data, dataIndices, ii, i1, leafNodeIndices, responses_); 
     return; 
    } 
    }

每次調用遞歸函數依賴於node.Feature.GetResponse()功能不同的計算時間。如果我使所有遞歸調用的計算時間相同（通過更改GetResponse()），代碼將平穩運行。

float AxisAlignedFeatureResponse::GetResponse(const IDataPointCollection& data, int index) const { 

    double retArg; 
    // retrieve DataManager object 
    const DataManager& concreteData = (const DataManager&)(data); 

//  // retrieve data point at index 
    DataPoint currDataPoint   = concreteData.getDataPoint(index); 
//  
//  // get coordinates of data point 
    Coordinate currCoordinates  = currDataPoint.getOrigPos(); 
//  
//  // get intensity image of the respective data point 
    int imgIndex = currDataPoint.getImageIndex(); 
    Image currImg = concreteData.getImage(imgIndex);  
    Image currFeatureImg = concreteData.getFeatureImage(imgIndex); 
    // return respective feature 
    int featureNumber = (int)(this->axis*(double)concreteData.getNumberOfFeatures());  
    if(featureNumber>=concreteData.getNumberOfFeatures()){ 
     cout<<"warning! trying to reach a feature that is not there!"<<endl; 
     featureNumber=concreteData.getNumberOfFeatures()-1; 
    } 

    std::vector<Coordinate> feature = concreteData.getFeature(featureNumber); 
    Coordinate tmp=currCoordinates+feature[0]; 
    if(feature[1].x == 0) { 
     retArg = currCoordinates.x*feature[0].x+currCoordinates.y*feature[0].y+currCoordinates.z*feature[0].z; 
    //retArg = 0;  //DOING THIS runs the code smoothly 
    } 

    else if(feature[1].x == 2) { 
     retArg = currFeatureImg.getValue(feature[0]); 
    } else { 
     retArg = currImg.mean(tmp,feature[1]); 
    } 
    return (float)(retArg); 
     //return (float) 0; 
}

來源

2016-03-08 Azhar

所以看起來可能問題出在GetResponse？代碼在哪裏？ – xaxxon

@xaxxon現在增加了，雖然我不明白除了計算時間差異之外怎麼會有問題 – Azhar

這看起來像valgrind的工作。除非你的程序突然終止，否則我看不到任何內存泄漏的原因。內存泄漏通常是由堆變量未被釋放引起的（「new」後面沒有後面的「delete」）。

Valgrind會給你一個罪魁禍首和行號，你可以設置斷點並通過gdb來查看究竟是怎麼回事。也許發佈運行「valgrind -v your_program」的詳細輸出。不要忘記使用-g選項進行編譯，以將完整的調試數據提供給valgrind和gdb。

來源

2016-03-08 06:04:28

不幸的是，沒有記錄的方式來運行valgrind和matlab mex文件 – Azhar

內存泄漏可能是由遞歸函數造成的

回答

相關問題