feat: use DTO for NCNN init parameters

jolibrain · Feb 12, 2021 · 635e2c1 · 635e2c1
1 parent 00036b1
commit 635e2c1
Show file tree

Hide file tree

Showing 2 changed files with 24 additions and 46 deletions.
diff --git a/src/backends/ncnn/ncnnlib.cc b/src/backends/ncnn/ncnnlib.cc
@@ -22,7 +22,6 @@
 #include "outputconnectorstrategy.h"
 #include <thread>
 #include <algorithm>
-#include "utils/utils.hpp"
 
 // NCNN
 #include "ncnnlib.h"
@@ -53,10 +52,10 @@ namespace dd
   {
     this->_libname = "ncnn";
     _net = new ncnn::Net();
-    _net->opt.num_threads = _threads;
+    _net->opt.num_threads = 1;
     _net->opt.blob_allocator = &_blob_pool_allocator;
     _net->opt.workspace_allocator = &_workspace_pool_allocator;
-    _net->opt.lightmode = _lightmode;
+    _net->opt.lightmode = true;
   }
 
   template <class TInputConnectorStrategy, class TOutputConnectorStrategy,
@@ -69,12 +68,9 @@ namespace dd
     this->_libname = "ncnn";
     _net = tl._net;
     tl._net = nullptr;
-    _nclasses = tl._nclasses;
-    _threads = tl._threads;
     _timeserie = tl._timeserie;
     _old_height = tl._old_height;
-    _inputBlob = tl._inputBlob;
-    _outputBlob = tl._outputBlob;
+    _init_dto = tl._init_dto;
   }
 
   template <class TInputConnectorStrategy, class TOutputConnectorStrategy,
@@ -94,6 +90,8 @@ namespace dd
   void NCNNLib<TInputConnectorStrategy, TOutputConnectorStrategy,
                TMLModel>::init_mllib(const APIData &ad)
   {
+    _init_dto = ad.createSharedDTO<NcnnInitDto>();
+
     bool use_fp32 = (ad.has("datatype")
                      && ad.get("datatype").get<std::string>()
                             == "fp32"); // default is fp16
@@ -124,35 +122,11 @@ namespace dd
     _old_height = this->_inputc.height();
     _net->set_input_h(_old_height);
 
-    if (ad.has("nclasses"))
-      _nclasses = ad.get("nclasses").get<int>();
-
-    if (ad.has("threads"))
-      _threads = ad.get("threads").get<int>();
-    else
-      _threads = dd_utils::my_hardware_concurrency();
-
     _timeserie = this->_inputc._timeserie;
     if (_timeserie)
       this->_mltype = "timeserie";
 
-    if (ad.has("lightmode"))
-      {
-        _lightmode = ad.get("lightmode").get<bool>();
-        _net->opt.lightmode = _lightmode;
-      }
-
-    // setting the value of Input Layer
-    if (ad.has("inputblob"))
-      {
-        _inputBlob = ad.get("inputblob").get<std::string>();
-      }
-    // setting the final Output Layer
-    if (ad.has("outputblob"))
-      {
-        _outputBlob = ad.get("outputblob").get<std::string>();
-      }
-
+    _net->opt.lightmode = _init_dto->lightmode;
     _blob_pool_allocator.set_size_compare_ratio(0.0f);
     _workspace_pool_allocator.set_size_compare_ratio(0.5f);
     model_type(this->_mlmodel._params, this->_mltype);
@@ -233,7 +207,10 @@ namespace dd
 
     // Extract detection or classification
     int ret = 0;
-    std::string out_blob = _outputBlob;
+    std::string out_blob;
+    if (_init_dto->outputBlob)
+      out_blob = _init_dto->outputBlob->std_str();
+
     if (out_blob.empty())
       {
         if (bbox == true)
@@ -262,11 +239,11 @@ namespace dd
       {
         best = ad_output.get("best").get<int>();
       }
-    if (best == -1 || best > _nclasses)
-      best = _nclasses;
+    if (best == -1 || best > _init_dto->nclasses)
+      best = _init_dto->nclasses;
 
       // for loop around batch size
-#pragma omp parallel for num_threads(_threads)
+#pragma omp parallel for num_threads(*_init_dto->threads)
     for (size_t b = 0; b < inputc._ids.size(); b++)
       {
         std::vector<double> probs;
@@ -276,8 +253,8 @@ namespace dd
         APIData rad;
 
         ncnn::Extractor ex = _net->create_extractor();
-        ex.set_num_threads(_threads);
-        ex.input(_inputBlob.c_str(), inputc._in.at(b));
+        ex.set_num_threads(_init_dto->threads);
+        ex.input(_init_dto->inputBlob->c_str(), inputc._in.at(b));
 
         ret = ex.extract(out_blob.c_str(), inputc._out.at(b));
         if (ret == -1)
@@ -423,7 +400,8 @@ namespace dd
       } // end for batch_size
 
     tout.add_results(vrad);
-    out.add("nclasses", this->_nclasses);
+    int nclasses = this->_init_dto->nclasses;
+    out.add("nclasses", nclasses);
     if (bbox == true)
       out.add("bbox", true);
     out.add("roi", false);

diff --git a/src/backends/ncnn/ncnnlib.h b/src/backends/ncnn/ncnnlib.h
@@ -22,12 +22,15 @@
 #ifndef NCNNLIB_H
 #define NCNNLIB_H
 
+#include "apidata.h"
+#include "utils/utils.hpp"
+
+#include "dto/ncnn.hpp"
+
 // NCNN
 #include "net.h"
 #include "ncnnmodel.h"
 
-#include "apidata.h"
-
 namespace dd
 {
   template <class TInputConnectorStrategy, class TOutputConnectorStrategy,
@@ -53,20 +56,17 @@ namespace dd
 
   public:
     ncnn::Net *_net = nullptr;
-    int _nclasses = 0;
     bool _timeserie = false;
-    bool _lightmode = true;
 
   private:
+    std::shared_ptr<NcnnInitDto> _init_dto;
     static ncnn::UnlockedPoolAllocator _blob_pool_allocator;
     static ncnn::PoolAllocator _workspace_pool_allocator;
 
   protected:
-    int _threads = 1;
     int _old_height = -1;
-    std::string _inputBlob = "data";
-    std::string _outputBlob;
   };
+
 }
 
 #endif