[Backend] Support the deployment of models which is larger than 2G in ORT and TRT backend (#514)

* support model bigger than 2G

* update code

* update code

* fix bug

* update code

cmake/paddle2onnx.cmake

@@ -43,7 +43,7 @@ else()
 endif(WIN32)
 
 set(PADDLE2ONNX_URL_BASE "https://bj.bcebos.com/fastdeploy/third_libs/")
-set(PADDLE2ONNX_VERSION "1.0.2rc")
+set(PADDLE2ONNX_VERSION "1.0.2")
 if(WIN32)
   set(PADDLE2ONNX_FILE "paddle2onnx-win-x64-${PADDLE2ONNX_VERSION}.zip")
   if(NOT CMAKE_CL_64)

fastdeploy/backends/ort/ort_backend.cc

@@ -82,7 +82,7 @@ bool OrtBackend::InitFromPaddle(const std::string& model_file,
   }
   char* model_content_ptr;
   int model_content_size = 0;
-
+  bool save_external = false;
 #ifdef ENABLE_PADDLE_FRONTEND
   paddle2onnx::CustomOp op;
   strcpy(op.op_name, "multiclass_nms3");
@@ -91,7 +91,7 @@ bool OrtBackend::InitFromPaddle(const std::string& model_file,
   if (!paddle2onnx::Export(model_file.c_str(), params_file.c_str(),
                            &model_content_ptr, &model_content_size, 11, true,
                            verbose, true, true, true, &op,
-                           1)) {
+                           1, "onnxruntime", nullptr, 0, "", &save_external)) {
     FDERROR << "Error occured while export PaddlePaddle to ONNX format."
             << std::endl;
     return false;
@@ -101,6 +101,15 @@ bool OrtBackend::InitFromPaddle(const std::string& model_file,
                                model_content_ptr + model_content_size);
   delete[] model_content_ptr;
   model_content_ptr = nullptr;
+  if(save_external){
+    std::string model_file_name = "model.onnx";
+    std::fstream f(model_file_name, std::ios::out);
+    FDASSERT(f.is_open(), "Can not open file: %s to save model.",
+                        model_file_name.c_str());
+    f << onnx_model_proto;
+    f.close();
+    return InitFromOnnx(model_file_name, option, false);
+  }
   return InitFromOnnx(onnx_model_proto, option, true);
 #else
   FDERROR << "Didn't compile with PaddlePaddle Frontend, you can try to "

fastdeploy/backends/tensorrt/trt_backend.cc

@@ -132,7 +132,7 @@ bool TrtBackend::InitFromPaddle(const std::string& model_file,
                            &model_content_ptr, &model_content_size, 11, true,
                            verbose, true, true, true, nullptr,
                            0, "tensorrt",
-                           &calibration_cache_ptr, &calibration_cache_size)) {
+                           &calibration_cache_ptr, &calibration_cache_size, "", &save_external_)) {
     FDERROR << "Error occured while export PaddlePaddle to ONNX format."
             << std::endl;
     return false;
@@ -148,6 +148,15 @@ bool TrtBackend::InitFromPaddle(const std::string& model_file,
     calibration_str_ = calibration_str;
     delete[] calibration_cache_ptr;
   }
+  if(save_external_){
+    model_file_name_ = "model.onnx";
+    std::fstream f(model_file_name_, std::ios::out);
+    FDASSERT(f.is_open(), "Can not open file: %s to save model.",
+                        model_file_name_.c_str());
+    f << onnx_model_proto;
+    f.close();
+    return InitFromOnnx(model_file_name_, option, false);
+  }
   return InitFromOnnx(onnx_model_proto, option, true);
 #else
   FDERROR << "Didn't compile with PaddlePaddle frontend, you can try to "
@@ -242,6 +251,10 @@ bool TrtBackend::InitFromOnnx(const std::string& model_file,
            "[ERROR] Error occurs while calling cudaStreamCreate().");
   }
 
+  if(save_external_){
+    onnx_content.clear();
+    onnx_content = model_file_name_;
+  }
   if (!CreateTrtEngineFromOnnx(onnx_content)) {
     FDERROR << "Failed to create tensorrt engine." << std::endl;
     return false;
@@ -593,7 +606,13 @@ bool TrtBackend::CreateTrtEngineFromOnnx(const std::string& onnx_model_buffer) {
     FDERROR << "Failed to call createParser()." << std::endl;
     return false;
   }
-  if (!parser_->parse(onnx_model_buffer.data(), onnx_model_buffer.size())) {
+  bool model_parser;
+  if(save_external_){
+    model_parser=!parser_->parseFromFile(onnx_model_buffer.c_str(), 0);
+  }else{
+    model_parser = !parser_->parse(onnx_model_buffer.data(), onnx_model_buffer.size());
+  }
+  if (model_parser) {
     FDERROR << "Failed to parse ONNX model by TensorRT." << std::endl;
     return false;
   }

fastdeploy/backends/tensorrt/trt_backend.h

@@ -130,6 +130,8 @@ class TrtBackend : public BaseBackend {
   std::map<std::string, int> io_name_index_;
 
   std::string calibration_str_;
+  bool save_external_ = false;
+  std::string model_file_name_ = "";
 
   // Sometimes while the number of outputs > 1
   // the output order of tensorrt may not be same