apollo::prediction::MultiAgentVehicleCpuTorch类 参考

#include <multi_agent_vehicle_torch_model.h>

类 apollo::prediction::MultiAgentVehicleCpuTorch 继承关系图:

apollo::prediction::MultiAgentVehicleCpuTorch 的协作图:

Public 成员函数
	MultiAgentVehicleCpuTorch ()
	Construct a new Multi Agent Vehicle Cpu Torch object
	~MultiAgentVehicleCpuTorch ()
	Destroy the Multi Agent Vehicle Cpu Torch object
virtual bool	Init ()
	parse model description class and load the model
virtual bool	Inference (const std::vector< void * > &input_buffer, unsigned int input_size, std::vector< void * > *output_buffer, unsigned int output_size)
	performing network inference
virtual bool	LoadModel ()
	load the model from file
virtual void	Destory ()
	free all memory requested, gpu or cpu
Public 成员函数 继承自 apollo::prediction::ModelBase
	ModelBase ()
	~ModelBase ()

额外继承的成员函数
Public 属性 继承自 apollo::prediction::ModelBase
std::string	model_path_
uint8_t	init_ = 0

详细描述

在文件 multi_agent_vehicle_torch_model.h 第 30 行定义.

构造及析构函数说明

MultiAgentVehicleCpuTorch()

apollo::prediction::MultiAgentVehicleCpuTorch::MultiAgentVehicleCpuTorch ( )

inline

Construct a new Multi Agent Vehicle Cpu Torch object

在文件 multi_agent_vehicle_torch_model.h 第 36 行定义.

{}

~MultiAgentVehicleCpuTorch()

apollo::prediction::MultiAgentVehicleCpuTorch::~MultiAgentVehicleCpuTorch ( )

inline

Destroy the Multi Agent Vehicle Cpu Torch object

在文件 multi_agent_vehicle_torch_model.h 第 41 行定义.

{Destory();}

成员函数说明

Destory()

void apollo::prediction::MultiAgentVehicleCpuTorch::Destory ( )

virtual

free all memory requested, gpu or cpu

返回

memory release result, true for success

实现了 apollo::prediction::ModelBase.

在文件 multi_agent_vehicle_torch_model.cc 第 147 行定义.

{}

Inference()

bool apollo::prediction::MultiAgentVehicleCpuTorch::Inference ( const std::vector< void * > &  input_buffer, unsigned int  input_size, std::vector< void * > *  output_buffer, unsigned int  output_size  )

virtual

performing network inference

参数

input_buffer	vector of input tensor
input_size	size of input_buffer
output_buffer	vector of output tensor
output_size	size of output_buffer

返回

init result, true for success

实现了 apollo::prediction::ModelBase.

在文件 multi_agent_vehicle_torch_model.cc 第 96 行定义.

                                                               {
  ACHECK(input_size == input_buffer.size() && input_size == 8);
  ACHECK(output_size == output_buffer->size() && output_size == 1);
 
  if (init_ == 0) {
    Init();
  }
 
  auto device = torch::Device(torch::kCPU);
  if (torch::cuda::is_available()) {
    device = torch::Device(torch::kCUDA);
  }
  torch::Tensor target_obstacle_pos =
      torch::from_blob(input_buffer[0], {1, max_agent_num, 20, 2});
  torch::Tensor target_obstacle_pos_step =
      torch::from_blob(input_buffer[1], {1, max_agent_num, 20, 2});
  torch::Tensor vector_data =
      torch::from_blob(input_buffer[2], {1, 450, 50, 9});
  auto options = torch::TensorOptions().dtype(torch::kBool);
  torch::Tensor vector_mask =
      torch::from_blob(input_buffer[3], {1, 450, 50}, options);
  torch::Tensor polyline_mask =
      torch::from_blob(input_buffer[4], {1, 450}, options);
  torch::Tensor rand_mask =
      torch::from_blob(input_buffer[5], {1, 450}, options);
  torch::Tensor polyline_id =
      torch::from_blob(input_buffer[6], {1, 450, 2});
  torch::Tensor obs_position =
      torch::from_blob(input_buffer[7], {1, max_agent_num, 3});
 
  std::vector<torch::jit::IValue> torch_inputs;
 
  torch_inputs.push_back(c10::ivalue::Tuple::create(
      {std::move(target_obstacle_pos.to(device)),
       std::move(target_obstacle_pos_step.to(device)),
       std::move(vector_data.to(device)),
       std::move(vector_mask.to(device)),
       std::move(polyline_mask.to(device)),
       std::move(rand_mask.to(device)),
       std::move(polyline_id.to(device)),
       std::move(obs_position.to(device))}));
 
  torch::Tensor torch_output_tensor =
      model_instance_.forward(torch_inputs).toTensor().to(torch::kCPU);
  memcpy((*output_buffer)[0], torch_output_tensor.data_ptr<float>(),
         1 * max_agent_num * 30 * 2 * sizeof(float));
  return true;
}

Init()

bool apollo::prediction::MultiAgentVehicleCpuTorch::Init ( )

virtual

parse model description class and load the model

参数

config_path

model config path

返回

init result, true for success

实现了 apollo::prediction::ModelBase.

在文件 multi_agent_vehicle_torch_model.cc 第 31 行定义.

                                     {
  ModelConf model_config;
  int status;
 
  if (init_ != 0) {
    return true;
  }
 
  std::string class_name =
      abi::__cxa_demangle(typeid(*this).name(), 0, 0, &status);
 
  std::string default_config_path =
      apollo::cyber::plugin_manager::PluginManager::Instance()
          ->GetPluginConfPath<ModelBase>(class_name,
                                         "conf/default_conf.pb.txt");
 
  if (!cyber::common::GetProtoFromFile(default_config_path, &model_config)) {
    AERROR << "Unable to load model conf file: " << default_config_path;
    return false;
  }
  model_path_ = model_config.model_path();
  init_ = 1;
 
  return LoadModel();
}

LoadModel()

bool apollo::prediction::MultiAgentVehicleCpuTorch::LoadModel ( )

virtual

load the model from file

返回

loading result, true for success

实现了 apollo::prediction::ModelBase.

在文件 multi_agent_vehicle_torch_model.cc 第 57 行定义.

                                          {
  auto device = torch::Device(torch::kCPU);
  if (torch::cuda::is_available()) {
    device = torch::Device(torch::kCUDA);
  }
 
  model_instance_ = torch::jit::load(model_path_, device);
 
  torch::set_num_threads(1);
 
  // Fake intput for the first frame
  torch::Tensor target_obstacle_pos =
    torch::randn({1, max_agent_num, 20, 2});
  torch::Tensor target_obstacle_pos_step =
    torch::randn({1, max_agent_num, 20, 2});
  torch::Tensor vector_data = torch::randn({1, 450, 50, 9});
  torch::Tensor vector_mask = torch::randn({1, 450, 50}) > 0.9;
  torch::Tensor polyline_mask = torch::randn({1, 450}) > 0.9;
  torch::Tensor rand_mask = torch::zeros({450}).toType(at::kBool);
  torch::Tensor polyline_id = torch::randn({1, 450, 2});
  torch::Tensor obs_position = torch::randn({1, max_agent_num, 3});
  std::vector<torch::jit::IValue> torch_inputs;
  torch::Tensor torch_default_output_tensor;
 
  torch_inputs.push_back(c10::ivalue::Tuple::create(
      {std::move(target_obstacle_pos.to(device)),
       std::move(target_obstacle_pos_step.to(device)),
       std::move(vector_data.to(device)),
       std::move(vector_mask.to(device)),
       std::move(polyline_mask.to(device)),
       std::move(rand_mask.to(device)),
       std::move(polyline_id.to(device)),
       std::move(obs_position.to(device))}));
 
  // warm up to avoid very slow first inference later
  WarmUp(torch_inputs, &model_instance_, &torch_default_output_tensor);
  return true;
}

---

该类的文档由以下文件生成:

• modules/prediction/evaluator/model_manager/model/multi_agent_vehicle_torch_cpu/multi_agent_vehicle_torch_model.h
• modules/prediction/evaluator/model_manager/model/multi_agent_vehicle_torch_cpu/multi_agent_vehicle_torch_model.cc