add part_load col

src/data.h

@@ -138,9 +138,10 @@ class IFMatrix {
   virtual utils::IIterator<ColBatch> *ColIterator(const std::vector<bst_uint> &fset) = 0;
   /*!
    * \brief check if column access is supported, if not, initialize column access
+   * \param enabled whether certain feature should be included in column access
    * \param subsample subsample ratio when generating column access
    */
-  virtual void InitColAccess(float subsample) = 0;
+  virtual void InitColAccess(const std::vector<bool> &enabled, float subsample) = 0;
   // the following are column meta data, should be able to answer them fast
   /*! \return whether column access is enabled */
   virtual bool HaveColAccess(void) const = 0;

src/io/page_fmatrix-inl.hpp

@@ -247,7 +247,7 @@ class FMatrixPage : public IFMatrix {
     size_t nmiss = buffered_rowset_.size() - (col_ptr[cidx+1] - col_ptr[cidx]);
     return 1.0f - (static_cast<float>(nmiss)) / buffered_rowset_.size();
   }
-  virtual void InitColAccess(float pkeep = 1.0f) {
+  virtual void InitColAccess(const std::vector<bool> &enabled, float pkeep = 1.0f) {
     if (this->HaveColAccess()) return;
     utils::Printf("start to initialize page col access\n");
     if (this->LoadColData()) {

src/io/simple_fmatrix-inl.hpp

@@ -48,9 +48,10 @@ class FMatrixS : public IFMatrix{
     size_t nmiss = buffered_rowset_.size() - (col_ptr_[cidx+1] - col_ptr_[cidx]);
     return 1.0f - (static_cast<float>(nmiss)) / buffered_rowset_.size();
   }
-  virtual void InitColAccess(float pkeep = 1.0f) {
+  virtual void InitColAccess(const std::vector<bool> &enabled, 
+                             float pkeep = 1.0f) {
     if (this->HaveColAccess()) return;
-    this->InitColData(pkeep);
+    this->InitColData(pkeep, enabled);
   }
   /*!
    * \brief get the row iterator associated with FMatrix
@@ -141,7 +142,7 @@ class FMatrixS : public IFMatrix{
    * \brief intialize column data
    * \param pkeep probability to keep a row
    */
-  inline void InitColData(float pkeep) {
+  inline void InitColData(float pkeep, const std::vector<bool> &enabled) {
     buffered_rowset_.clear();
     // note: this part of code is serial, todo, parallelize this transformer
     utils::SparseCSRMBuilder<RowBatch::Entry> builder(col_ptr_, col_data_);
@@ -155,11 +156,13 @@ class FMatrixS : public IFMatrix{
           buffered_rowset_.push_back(static_cast<bst_uint>(batch.base_rowid+i));
           RowBatch::Inst inst = batch[i];
           for (bst_uint j = 0; j < inst.length; ++j) {
+            if (enabled[inst[j].index]){ 
               builder.AddBudget(inst[j].index);
             }
           }
         }
       }
+    }
     builder.InitStorage();
 
     iter_->BeforeFirst();
@@ -172,6 +175,7 @@ class FMatrixS : public IFMatrix{
           ++ktop;
           RowBatch::Inst inst = batch[i];
           for (bst_uint j = 0; j < inst.length; ++j) {
+            if (enabled[inst[j].index]) { 
               builder.PushElem(inst[j].index,
                                Entry((bst_uint)(batch.base_rowid+i),
                                      inst[j].fvalue));
@@ -179,6 +183,7 @@ class FMatrixS : public IFMatrix{
           }
         }
       }
+    }
     // sort columns
     bst_omp_uint ncol = static_cast<bst_omp_uint>(this->NumCol());
     #pragma omp parallel for schedule(static)

src/learner/learner-inl.hpp

@@ -31,6 +31,7 @@ class BoostLearner {
     name_gbm_ = "gbtree";
     silent= 0;
     prob_buffer_row = 1.0f;
+    part_load_col = 0;
   }
   ~BoostLearner(void) {
     if (obj_ != NULL) delete obj_;
@@ -88,6 +89,7 @@ class BoostLearner {
       this->SetParam(n.c_str(), val);
     }
     if (!strcmp(name, "silent")) silent = atoi(val);
+    if (!strcmp(name, "part_load_col")) part_load_col = atoi(val);
     if (!strcmp(name, "prob_buffer_row")) {
       prob_buffer_row = static_cast<float>(atof(val));
       this->SetParam("updater", "grow_colmaker,refresh,prune");
@@ -165,7 +167,40 @@ class BoostLearner {
    * \param p_train pointer to the matrix used by training
    */
   inline void CheckInit(DMatrix *p_train) {
-    p_train->fmat()->InitColAccess(prob_buffer_row);
+    int ncol = p_train->info.info.num_col;    
+    std::vector<bool> enabled(ncol, true);
+    
+    if (part_load_col != 0) {      
+      std::vector<unsigned> col_index;
+      for (int i = 0; i < ncol; ++i) {
+        col_index.push_back(i);
+      }
+      random::Shuffle(col_index);
+      std::string s_model;
+      utils::MemoryBufferStream ms(&s_model);
+      utils::IStream &fs = ms;
+      if (sync::GetRank() == 0) {
+        fs.Write(col_index);
+        sync::Bcast(&s_model, 0);
+      } else {
+        sync::Bcast(&s_model, 0);
+        fs.Read(&col_index);
+      }
+      int nsize = sync::GetWorldSize();
+      int step = (ncol + nsize -1) / nsize;
+      int pid = sync::GetRank();
+      std::fill(enabled.begin(), enabled.end(), false);
+      int start = step * pid;
+      int end = std::min(step * (pid + 1), ncol);
+      utils::Printf("rank %d idset:", pid);
+      for (int i = start; i < end; ++i) {
+        enabled[col_index[i]] = true;
+        utils::Printf(" %u", col_index[i]);
+      }
+      utils::Printf("\n");
+    }
+    // initialize column access
+    p_train->fmat()->InitColAccess(enabled, prob_buffer_row);    
   }
   /*!
    * \brief update the model for one iteration
@@ -316,6 +351,8 @@ class BoostLearner {
   // data fields
   // silent during training
   int silent;
+  // randomly load part of data
+  int part_load_col;
   // maximum buffred row value
   float prob_buffer_row;
   // evaluation set

src/sync/sync.h

@@ -21,6 +21,9 @@ enum ReduceOp {
 
 /*! \brief get rank of current process */
 int GetRank(void);
+/*! \brief get total number of process */
+int GetWorldSize(void);
+
 /*! 
  * \brief this is used to check if sync module is a true distributed implementation, or simply a dummpy
  */

src/sync/sync_empty.cpp

@@ -17,6 +17,10 @@ bool IsDistributed(void) {
   return false;
 }
 
+int GetWorldSize(void) {
+  return 1;
+}
+
 template<>
 void AllReduce<uint32_t>(uint32_t *sendrecvbuf, int count, ReduceOp op) {
 }

src/sync/sync_mpi.cpp

@@ -8,6 +8,10 @@ int GetRank(void) {
   return MPI::COMM_WORLD.Get_rank();
 }
 
+int GetWorldSize(void) {
+  return MPI::COMM_WORLD.Get_size();
+}
+
 void Init(int argc, char *argv[]) {
   MPI::Init(argc, argv);
 }

src/xgboost_main.cpp

@@ -160,7 +160,7 @@ class BoostLearnTask {
       if (!silent) printf("boosting round %d, %lu sec elapsed\n", i, elapsed);
       learner.UpdateOneIter(i, *data); 
       std::string res = learner.EvalOneIter(i, devalall, eval_data_names);
-      if (silent < 1) {
+      if (silent < 2) {
         fprintf(stderr, "%s\n", res.c_str());
       }
       if (save_period != 0 && (i + 1) % save_period == 0) {