FeatureOutput 学习数据输出
源码位置：
modules/planning/planning_base/common/feature_output.h/.cc
模块定位
FeatureOutput 是一个静态工具类，负责将规划过程中的学习数据帧（LearningDataFrame）收集并序列化到磁盘文件。用于离线训练数据采集，支持学习型规划器的数据管道。
类声明
cpp
class FeatureOutput {
public:
FeatureOutput() = delete; // 纯静态类
static void Close();
static void Clear();
static bool Ready();
static void InsertLearningDataFrame(const std::string& record_filename,
const LearningDataFrame& learning_data_frame);
static void InsertPlanningResult();
static LearningDataFrame* GetLatestLearningDataFrame();
static void WriteLearningData(const std::string& record_file);
static void WriteRemainderiLearningData(const std::string& record_file);
static int SizeOfLearningData();
private:
static LearningData learning_data_;
static int learning_data_file_index_;
};方法详解
Clear() / Close()
cpp
void FeatureOutput::Clear() {
learning_data_.Clear();
learning_data_file_index_ = 0;
}
void FeatureOutput::Close() { Clear(); }- 重置内部状态，清空累积的学习数据
feature_output.cc:L30-L35
Ready()
cpp
bool FeatureOutput::Ready() {
Clear();
return true;
}- 清空后返回就绪状态
feature_output.cc:L37-L40
InsertLearningDataFrame()
cpp
void FeatureOutput::InsertLearningDataFrame(
const std::string& record_file,
const LearningDataFrame& learning_data_frame) {
learning_data_.add_learning_data_frame()->CopyFrom(learning_data_frame);
if (learning_data_.learning_data_frame_size() >=
FLAGS_learning_data_frame_num_per_file) {
WriteLearningData(record_file);
}
}- 将一帧学习数据追加到内部缓冲
- 当帧数达到
FLAGS_learning_data_frame_num_per_file时自动写入文件 feature_output.cc:L42-L52
GetLatestLearningDataFrame()
cpp
LearningDataFrame* FeatureOutput::GetLatestLearningDataFrame() {
const int size = learning_data_.learning_data_frame_size();
return size > 0 ? learning_data_.mutable_learning_data_frame(size - 1) : nullptr;
}- 返回最新一帧的可变指针，用于后续补充数据
feature_output.cc:L54-L58
WriteLearningData()
cpp
void FeatureOutput::WriteLearningData(const std::string& record_file) {
std::string src_file_name = record_file.substr(record_file.find_last_of("/") + 1);
src_file_name = src_file_name.empty() ? "00000" : src_file_name;
const std::string dest_file = absl::StrCat(
FLAGS_planning_data_dir, "/", src_file_name, ".",
learning_data_file_index_, ".bin");
cyber::common::SetProtoToBinaryFile(learning_data_, dest_file);
learning_data_.Clear();
learning_data_file_index_++;
}- 输出路径：
{FLAGS_planning_data_dir}/{record_filename}.{index}.bin - 使用 protobuf 二进制格式序列化
- 写入后清空缓冲并递增文件索引
feature_output.cc:L62-L73
WriteRemainderiLearningData()
cpp
void FeatureOutput::WriteRemainderiLearningData(const std::string& record_file) {
if (learning_data_.learning_data_frame_size() > 0) {
WriteLearningData(record_file);
}
}- 将缓冲中剩余的不足一批的数据强制写出
- 通常在录制结束时调用
feature_output.cc:L75-L80
配置项
| 参数 | 说明 |
|---|---|
FLAGS_learning_data_frame_num_per_file | 每个文件包含的帧数 |
FLAGS_planning_data_dir | 学习数据输出目录 |
数据流
PlanningComponent::RunOnce()
└── LearningModelInferenceTask / FeatureGenerator
└── FeatureOutput::InsertLearningDataFrame()
└── 累积到阈值 → WriteLearningData() → .bin 文件
Steven Moder