caffe中的工厂模式

1. 定义一个名字（字符串）与函数一一映射关系，通过REGISTER_LAYER_CREATOR将名字以及对应的函数注册到系统中

//define a function
shared_ptr<LayerBase> GetConvolutionLayer(const LayerParameter& param,
    Type ftype, Type btype, size_t) {
    //do something
    }
    shared_ptr<LayerBase> GetSigmoidLayer(const LayerParameter& param,
    Type ftype, Type btype, size_t){}
shared_ptr<LayerBase> GetSoftmaxLayer(const LayerParameter& param,
    Type ftype, Type btype, size_t)  {}
 //regist "Convolution" to GetConvolutionLayer function
REGISTER_LAYER_CREATOR(Convolution, GetConvolutionLayer);
//实际上会扩展为：static LayerRegisterer g_creator_Convolution("Convolution", GetConvolutionLayer);
//其中REGISTER_LAYER_CREATOR宏定义如下：
#define REGISTER_LAYER_CREATOR(type, creator)                                  \
  static LayerRegisterer g_creator_##type(#type, creator);

2.　通过g_registry_来保存键值对

class LayerRegistry {
 public:
  typedef shared_ptr<LayerBase> (*Creator)(const LayerParameter&, Type, Type, size_t);
  typedef std::map<string, Creator> CreatorRegistry;

  static CreatorRegistry& Registry() {
    static CreatorRegistry g_registry_;
    return g_registry_;
  }

2. 最后就可以通过create_layer接口来创建新的layer

  static shared_ptr<LayerBase> CreateLayer(const LayerParameter& param, size_t solver_rank) {
  //...
      return registry[layer_type](param, ftype, btype, solver_rank);
  }

更详细内容需要参考layer_factory.cpp文件．