batching if huge number of descriptors

imagerepresentation.cpp

@@ -7,11 +7,13 @@
 #ifdef _OPENMP
 #include <omp.h>
 #endif
-
+#include <algorithm>
+#include <iterator>
 
 #define VERBOSE 1
 #include <zmq.hpp>
 
+
 inline static bool endsWith(const std::string& str, const std::string& suffix)
 {
   return str.size() >= suffix.size() && 0 == str.compare(str.size()-suffix.size(), suffix.size(), suffix);
@@ -20,51 +22,83 @@ std::vector<std::vector<float> > DescribeWithZmq(zmqDescriptorParams par,
                                                  AffineRegionVector &kps,
                                                  SynthImage &temp_img1){
 
+  int kp_size = kps.size() ;
   std::vector<std::vector<float> > out;
-  cv::Mat patches;
-  int odd_patch_size = par.patchSize ;
-  if (kps.size() > 0){
-      ExtractPatchesColumn(kps,temp_img1, patches,
-                           par.mrSize,
-                           odd_patch_size,
-                           false,
-                           false,
-                           true);
-    }
+  int batch_size = 2000;
+  std::cout << kp_size << std::endl;
+  if (kp_size <= batch_size){
+
+
 
-  std::vector<uchar> bufff1;
-  cv::imencode(".png",patches,bufff1);
+      cv::Mat patches;
+      int odd_patch_size = par.patchSize ;
+      if (kps.size() > 0){
+          ExtractPatchesColumn(kps,temp_img1, patches,
+                               par.mrSize,
+                               odd_patch_size,
+                               false,
+                               false,
+                               true);
 
-  zmq::context_t context (1);
-  int socket_mode = ZMQ_REQ;
 
-  zmq::socket_t   socket(context, socket_mode);
-  socket.connect(par.port);
+          std::vector<uchar> bufff1;
+          cv::imencode(".png",patches,bufff1);
 
-  zmq::message_t mods_to_cnn(bufff1.size()) ;
-  memcpy ((void *) mods_to_cnn.data (), bufff1.data(), bufff1.size());
-  zmq::message_t cnn_to_mods;
+          zmq::context_t context (1);
+          int socket_mode = ZMQ_REQ;
+
+          zmq::socket_t   socket(context, socket_mode);
+          socket.connect(par.port);
+
+          zmq::message_t mods_to_cnn(bufff1.size()) ;
+          memcpy ((void *) mods_to_cnn.data (), bufff1.data(), bufff1.size());
+          zmq::message_t cnn_to_mods;
 #pragma omp critical
-  {
-    socket.send(mods_to_cnn);
-    //  Get the reply.
-    socket.recv(&cnn_to_mods);
-  }
-  std::vector<float> inMsg(cnn_to_mods.size() / sizeof(float));
-  std::memcpy(inMsg.data(), cnn_to_mods.data(), cnn_to_mods.size());
-  const int desc_size = inMsg.size() /kps.size() ;
-
-  for (int img_num=0; img_num<kps.size(); img_num++)
-    {
-      std::vector<float> curr_desc(desc_size);
-      int offset = img_num*desc_size;
-      for (int i = 0; i < desc_size; ++i) {
-          const float v1 = inMsg[i+offset];
-          curr_desc[i] = v1;
+          {
+            socket.send(mods_to_cnn);
+            //  Get the reply.
+            socket.recv(&cnn_to_mods);
+          }
+          std::vector<float> inMsg(cnn_to_mods.size() / sizeof(float));
+          std::memcpy(inMsg.data(), cnn_to_mods.data(), cnn_to_mods.size());
+          const int desc_size = inMsg.size() /kps.size() ;
+
+          for (int img_num=0; img_num<kps.size(); img_num++)
+            {
+              std::vector<float> curr_desc(desc_size);
+              int offset = img_num*desc_size;
+              for (int i = 0; i < desc_size; ++i) {
+                  const float v1 = inMsg[i+offset];
+                  curr_desc[i] = v1;
+                }
+              out.push_back(curr_desc);
+            }
+          socket.close();
+        }
+    } else {
+      int num_batches = (kp_size / batch_size) + 1;
+      AffineRegionVector::const_iterator startIter(kps.cbegin());
+      AffineRegionVector::const_iterator endIter(kps.cbegin());
+      int num_done = 0;
+      int to_add = 0;
+      for (int bi = 0; bi < num_batches; bi++) {
+          num_done += batch_size;
+          if (num_done > kp_size) {
+              num_done -= batch_size;
+              to_add = kp_size - num_done;
+            } else {
+              to_add = batch_size;
+            }
+          std::advance(endIter, to_add);
+
+          AffineRegionVector current_kps(startIter, endIter);
+          std::vector<std::vector<float> > current_out = DescribeWithZmq(par,
+                                                                         current_kps,
+                                                                         temp_img1);
+          std::move(current_out.begin(), current_out.end(), std::back_inserter(out));
+          std::advance(startIter, to_add);
         }
-      out.push_back(curr_desc);
     }
-  socket.close();
   return out;
 }
 
@@ -1318,7 +1352,7 @@ void ImageRepresentation::LoadRegions(std::string fname) {
     }
   kpfile.close();
 }
- AffineRegionVector ImageRepresentation::PreLoadRegionsNPZ(std::string fname) {
+AffineRegionVector ImageRepresentation::PreLoadRegionsNPZ(std::string fname) {
   cnpy::npz_t my_npz = cnpy::npz_load(fname);
   std::vector<std::string> keys;
   bool A_is_here = false;