Improved general usage and added CMake integration

.gitignore

@@ -0,0 +1,4 @@
+*~
+
+*.o
+*.swp

CMakeLists.txt

@@ -0,0 +1,10 @@
+CMAKE_MINIMUM_REQUIRED( VERSION 3.3.1 )
+PROJECT( MLGkernel CXX )
+
+SET( CMAKE_CXX_STANDARD 11 )
+SET( CMAKE_CXX_STANDARD_REQUIRED ON )
+
+FIND_PACKAGE( Eigen3 REQUIRED )
+FIND_PACKAGE( Threads REQUIRED )
+
+ADD_SUBDIRECTORY( MLGkernel )

MLGkernel/CMakeLists.txt

@@ -0,0 +1,19 @@
+FILE(GLOB SOURCES
+  *.cpp
+  ../utility/*.cpp
+  ../matrices/matrices.cpp
+)
+
+INCLUDE_DIRECTORIES(
+  ${CMAKE_SOURCE_DIR}/include
+  ${CMAKE_SOURCE_DIR}/matrices
+  ${CMAKE_SOURCE_DIR}/utility
+  ${CMAKE_SOURCE_DIR}/utility/filetypes
+)
+
+ADD_EXECUTABLE( runMLG
+  ${SOURCES}
+)
+
+TARGET_INCLUDE_DIRECTORIES( runMLG SYSTEM PUBLIC ${EIGEN3_INCLUDE_DIR} )
+TARGET_LINK_LIBRARIES( runMLG Threads::Threads )

MLGkernel/FLGkernel.cpp

@@ -35,10 +35,18 @@ double FLGkernel::operator()(const FLGinstance& x1, const FLGinstance& x2) const
 
   Cvector lambda=(x1.Sinv+x2.Sinv).eigenvalues();
   //double detS=1; for(int i=0; i<lambda.n; i++) detS*=lambda(i); detS=1.0/detS; 
-  double log_detS=0; for(int i=0; i<lambda.n; i++) log_detS-=log(lambda(i)); 
+  double log_detS=0;
+  for(int i=0; i<lambda.n; i++)
+  {
+    if( isfinite( lambda(i) ) )
+      log_detS-=log(lambda(i));
+  }
+
+  auto det_x1 = isfinite( x1.log_detS ) ? x1.log_detS : 0.0;
+  auto det_x2 = isfinite( x2.log_detS ) ? x2.log_detS : 0.0;
 
   //double r=sqrt(detS/sqrt(x1.detS*x2.detS));
-  double logr=(log_detS-0.5*(x1.log_detS+x2.log_detS))/2;
+  double logr=(log_detS-0.5*(det_x1 + det_x2))/2;
   double r=0;
   if(logr<-30){cout<<"Underflow!"<<endl;} 
   else r=exp(logr);

MLGkernel/MLGdataset.cpp

@@ -29,7 +29,10 @@
 #include "FLGkernel.hpp"
 #include "MatrixOF_ASCII.hpp"
 #include "params.hpp"
+
 #include <string>
+#include <unordered_map>
+#include <set>
 
 void MLGdataset::condense(const int nlevels, const int leaf_radius){
   assert(nlevels>0);
@@ -100,9 +103,16 @@ void MLGdataset::loadGraphs(std::string filename){
   int i=0;
   int n;
   ifs >> numGraphs;
-  while(ifs.good()){
+
+  graphs.reserve( numGraphs );
+
+  while( ifs ){
     ifs>>n;
-    if(!ifs.good()) break;
+
+    // done with reading; skip!
+    if( !ifs )
+      break;
+
     //cout<<"Reading graph "<<++i<<" (n="<<n<<")"<<endl;
     Cmatrix A(n,n);
     for(int j=0; j<n; j++){
@@ -114,27 +124,73 @@ void MLGdataset::loadGraphs(std::string filename){
   assert(numGraphs == graphs.size());
 }
 
-void MLGdataset::loadDiscreteFeatures(std::string filename, int numFeatures){
+void MLGdataset::loadDiscreteFeatures( std::string filename ) {
   ifstream ifs(filename);
-  if(ifs.fail()){
+
+  if( !ifs ) {
     cout << "Failed to open " << filename << "." << endl;
-    exit(0);
+    exit( -1 );
   }
+
   int numVertices = 0;
-  int graphIndex = 0;
   int label;
   int numGraphs;
   ifs >> numGraphs;
-  while(ifs.good()){
-    ifs>>numVertices;
-    if(!ifs.good()) break;
-    for(int i=0; i<numVertices; i++){
+
+  // Store the position of the stream so that we can roll it back later
+  // on *after* having counted the number of features.
+  auto position = ifs.tellg();
+
+  // Stores the labels that are encountered. Since this is a set, the
+  // label count is guaranteed to be unique.
+  set<int> labels;
+
+  while( ifs ) {
+    ifs >> numVertices; // ignore this for now as we do not need it for counting
+
+    for( int i = 0; i < numVertices; i++ ) {
+      ifs >> label;
+      labels.insert( label );
+    }
+  }
+
+  int numFeatures = static_cast<int>( labels.size() );
+  cout << "Number of features: " << numFeatures << endl;
+
+  // Create a mapping of labels to indices. This makes it possible to
+  // load and handle graphs that have non-contiguous label sequences.
+
+  unordered_map<int, int> label_to_index;
+
+  {
+    int index = 0;
+    for( auto&& label : labels )
+      label_to_index[label] = index++;
+  }
+
+  // Reset the stream and convert the labels now into their graph
+  // representation.
+  ifs.clear();
+  ifs.seekg( position );
+
+  int graphIndex = 0;
+
+  while( ifs ) {
+    ifs >> numVertices;
+
+    if( !ifs )
+      break;
+
+    for( int i = 0; i < numVertices; i++ ) {
       ifs >> label;
+
       graphs[graphIndex]->labels[i] = Cvector::Zero(numFeatures+1);
-      graphs[graphIndex]->labels[i](label) = 1;
+      graphs[graphIndex]->labels[i](label_to_index[label]) = 1;
     }
+
     graphIndex++;
   }
+
   assert(graphs.size() == numGraphs);
   assert(graphIndex == graphs.size());
 }

MLGkernel/MLGdataset.hpp

@@ -47,7 +47,7 @@ class MLGdataset{
 public:
 
   void loadGraphs(std::string filename);
-  void loadDiscreteFeatures(std::string filename, int numFeatures);
+  void loadDiscreteFeatures(std::string filename);
   void loadFeatures(std::string filename);
   void saveGram(std::string filename);
   void fillGram(double *npmatrix, int rows, int cols);

MLGkernel/runMLG.cpp

@@ -61,19 +61,6 @@ string genSaveName(string data, double eta, double gamma, int radius, int levels
   return ss.str();
 }
 
-// These are the number of discrete node labels for each of the benchmark datasets.
-int get_num_features(string features){
-  if(features.find("MUTAG") != string::npos) return 7;
-  if(features.find("PTC") != string::npos) return 22;
-  if(features.find("PROTEINS") != string::npos) return 3;
-  if(features.find("NCI109") != string::npos) return 38;
-  if(features.find("NCI1") != string::npos) return 37;
-
-  cout << "Supplied dataset is not one of the sample datasets! You can manually change this code to use the correct number of discrete features of your dataset." << endl;
-  exit(0);
-  return 0;
-}
-
 void runMLG(Params& p) {
   threadManager.maxthreads = p.num_threads;
   MLGdataset dataset(p.data_path, p.eta, p.gamma, p.grow_or_double);
@@ -83,10 +70,8 @@ void runMLG(Params& p) {
     cout << "Computing degree features" << endl;
     for(auto g: dataset.graphs) g->computeDegreeFeatures(20); // all sample datasets have max degree < 20
   } else {
-    cout << "Computing discrete features" << endl;
-    int num_features = get_num_features(p.features_path);
-    cout << "num features: " << num_features << endl;
-    dataset.loadDiscreteFeatures(p.features_path, num_features);
+    cout << "Loading discrete features" << endl;
+    dataset.loadDiscreteFeatures(p.features_path);
   }
 
   dataset.computeGram(p.levels, p.radius);

README.md

@@ -9,6 +9,30 @@ R. Kondor, H. Pan, [The Multiscale Graph Laplacian](https://arxiv.org/abs/1603.0
 * [Eigen](http://eigen.tuxfamily.org/index.php)
 
 ## Installation/Setup
+
+### Automated installation using `CMake`
+
+The project is configured to use `CMake` to provide a streamlined installation
+experience. After installing `CMake` and `eigen3` using your favourite package
+manager, the following commands are sufficient to compile the
+executable:
+
+```bash
+$ mkdir build
+$ cd build
+$ cmake ../
+$ make -j4
+```
+
+When using Mac OS X, we recommend the [Homebrew](https://brew.sh)
+package manager for installing the dependencies:
+
+```bash
+$ brew install cmake eigen3
+```
+
+### Manual installation using `make`
+
 Change the EIGENDIR variable Makefile.options to the path to your installation of
 the Eigen library. Run the following command to create the runMLG executable in the MLGkernel directory.
 ```bash

matrices/Cvector.hpp

@@ -116,8 +116,8 @@ class Cvector: public DenseVector, public Serializable{
 
 public: // element access 
 
-  FIELD& operator()(const int i){return array[i];}
-  FIELD operator()(const int i) const {return array[i];}
+  FIELD& operator()(const int i){ assert( i < n ); return array[i];}
+  FIELD operator()(const int i) const { assert( i < n ); return array[i];}
 
   void (foreach)(std::function<void(const INDEX, FIELD&)> lambda) {for(int i=0; i<n; i++) lambda(i,array[i]);}
   void (foreach)(std::function<void(const INDEX, const FIELD)> lambda) const {for(int i=0; i<n; i++) lambda(i,array[i]);}