mounty_fsc

（Caffe）基本类Blob，Layer，Net（一）

本文地址：http://blog.csdn.net/mounty_fsc/article/details/51085654

Caffe中，Blob，Layer，Net，Solver是最为核心的类，以下介绍这几个类，Solver将在下一节介绍。

1 Blob

1.1 简介

Blob是：

对待处理数据带一层封装，用于在Caffe中通信传递。
也为CPU和GPU间提供同步能力
数学上，是一个N维的C风格的存储数组
总的来说，Caffe使用Blob来交流数据，其是Caffe中标准的数组与统一的内存接口，它是多功能的，在不同的应用场景具有不同的含义，如可以是：batches of images, model parameters, and derivatives for optimization等。

1.2 源代码

/** * @brief A wrapper around SyncedMemory holders serving as the basic * computational unit through which Layer%s, Net%s, and Solver%s * interact. * * TODO(dox): more thorough description. */  
template <typename Dtype>  
class Blob {  
 public:  
  Blob()  
       : data_(), diff_(), count_(0), capacity_(0) {}  

  /// @brief Deprecated; use <code>Blob(const vector<int>& shape)</code>. 
  explicit Blob(const int num, const int channels, const int height,  
      const int width);  
  explicit Blob(const vector<int>& shape);  

  .....  

 protected:  
  shared_ptr<SyncedMemory> data_;  
  shared_ptr<SyncedMemory> diff_;  
  shared_ptr<SyncedMemory> shape_data_;  
  vector<int> shape_;  
  int count_;  
  int capacity_;  

  DISABLE_COPY_AND_ASSIGN(Blob);  
};  // class Blob

注：此处只保留了构造函数与成员变量。

说明：

Blob在实现上是对SyncedMemory（见1.4部分）进行了一层封装。
shape_为blob维度，见1.3部分
data_为原始数据
diff_为梯度信息
count为该blob的总容量（即数据的size），函数count(x,y)（或count(x)）返回某个切片[x,y]（[x,end]）内容量，本质上就是shape[x]shape[x+1]….*shape[y]的值

1.3 Blob的shape

由源代码中可以注意到Blob有个成员变量：vector shape_
其作用：

对于图像数据，shape可以定义为4维的数组(Num, Channels, Height, Width)或(n, k, h, w)，所以Blob数据维度为n*k*h*w，Blob是row-major保存的，因此在(n, k, h, w)位置的值物理位置为((n * K + k) * H + h) * W + w。其中Number是数据的batch size，对于256张图片为一个training batch的ImageNet来说n = 256;Channel是特征维度，如RGB图像k = 3
对于全连接网络，使用2D blobs (shape (N, D))，然后调用InnerProductLayer
对于参数，维度根据该层的类型和配置来确定。对于有3个输入96个输出的卷积层，Filter核 11 x 11，则blob为96 x 3 x 11 x 11. 对于全连接层，1000个输出，1024个输入，则blob为1000 x 1024.

1.4 SyncedMemory

由1.2知，Blob本质是对SyncedMemory的再封装。其核心代码如下：

/** 
 * @brief Manages memory allocation and synchronization between the host (CPU) 
 *        and device (GPU). 
 * 
 * TODO(dox): more thorough description. 
 */  
class SyncedMemory {  
 public:  
...  
 const void* cpu_data();  
  const void* gpu_data();  
  void* mutable_cpu_data();  
  void* mutable_gpu_data();  
...  
 private:  
...  
  void* cpu_ptr_;  
  void* gpu_ptr_;  
...  
};  // class SyncedMemory

Blob同时保存了data_和diff_,其类型为SyncedMemory的指针。
对于data_(diff_相同),其实际值要么存储在CPU(cpu_ptr_)要么存储在GPU(gpu_ptr_),有两种方式访问CPU数据(GPU相同)：

常量方式,void* cpu_data()，其不改变cpu_ptr_指向存储区域的值。

可变方式，void* mutable_cpu_data()，其可改变cpu_ptr_指向存储区值。
以mutable_cpu_data()为例

void* SyncedMemory::mutable_cpu_data() {
  to_cpu();
  head_ = HEAD_AT_CPU;
  return cpu_ptr_;
}

inline void SyncedMemory::to_cpu() {
  switch (head_) {
  case UNINITIALIZED:
    CaffeMallocHost(&cpu_ptr_, size_, &cpu_malloc_use_cuda_);
    caffe_memset(size_, 0, cpu_ptr_);
    head_ = HEAD_AT_CPU;
    own_cpu_data_ = true;
    break;
  case HEAD_AT_GPU:

#ifndef CPU_ONLY

    if (cpu_ptr_ == NULL) {
      CaffeMallocHost(&cpu_ptr_, size_, &cpu_malloc_use_cuda_);
      own_cpu_data_ = true;
    }
    caffe_gpu_memcpy(size_, gpu_ptr_, cpu_ptr_);
    head_ = SYNCED;

#else

    NO_GPU;

#endif

    break;
  case HEAD_AT_CPU:
  case SYNCED:
    break;
  }
}

说明：

经验上来说，如果不需要改变其值，则使用常量调用的方式，并且，不要在你对象中保存其指针。为何要这样设计呢，因为这样涉及能够隐藏CPU到GPU的同步细节，以及减少数据传递从而提高效率，当你调用它们的时候，SyncedMem会决定何时去复制数据，通常情况是仅当gnu或cpu修改后有复制操作，引用1官方文档中有一个例子说明何时进行复制操作。
调用mutable_cpu_data()可以让head转移到cpu上
第一次调用mutable_cpu_data()是UNINITIALIZED将执行9到14行，将为cpu_ptr_分配host内存
若head从gpu转移到cpu，将把数据从gpu复制到cpu中

2 Layer

2.1 简介

Layer是Caffe的基础以及基本计算单元。Caffe十分强调网络的层次性，可以说，一个网络的大部分功能都是以Layer的形式去展开的，如convolute,pooling,loss等等。
在创建一个Caffe模型的时候，也是以Layer为基础进行的，需按照src/caffe/proto/caffe.proto中定义的网络及参数格式定义网络 prototxt文件（需了解google protocol buffer）

2.2 Layer与Blob的关系

如图，名为conv1的Layer 的输入是名为data的bottom blob，其输出是名为conv1的top blob。

其protobuff定义如下，一个layer有一个到多个的top和bottom，其对应于blob

layer { name: "conv1" type: "Convolution" bottom: "data" top: "conv1" .... }

2.3 源代码

 /** * Layer%s must implement a Forward function, in which they take their input * (bottom) Blob%s (if any) and compute their output Blob%s (if any). * They may also implement a Backward function, in which they compute the error * gradients with respect to their input Blob%s, given the error gradients with * their output Blob%s. */  
    template <typename Dtype>  
    class Layer {  
     public:  
      /** * You should not implement your own constructor. Any set up code should go * to SetUp(), where the dimensions of the bottom blobs are provided to the * layer. */  
      explicit Layer(const LayerParameter& param)  
        : layer_param_(param), is_shared_(false) {  
    ...  
        }  
      virtual ~Layer() {}  

      /** * @brief Implements common layer setup functionality. * @param bottom the preshaped input blobs * @param top * the allocated but unshaped output blobs, to be shaped by Reshape */  
      void SetUp(const vector<Blob<Dtype>*>& bottom,  
          const vector<Blob<Dtype>*>& top) {  
    ...  
      }  

      ...  

      /** * @brief Given the bottom blobs, compute the top blobs and the loss. * \return The total loss from the layer. * * The Forward wrapper calls the relevant device wrapper function * (Forward_cpu or Forward_gpu) to compute the top blob values given the * bottom blobs. If the layer has any non-zero loss_weights, the wrapper * then computes and returns the loss. * * Your layer should implement Forward_cpu and (optionally) Forward_gpu. */  
      inline Dtype Forward(const vector<Blob<Dtype>*>& bottom,  
          const vector<Blob<Dtype>*>& top);  

      /** * @brief Given the top blob error gradients, compute the bottom blob error * gradients. * * @param top * the output blobs, whose diff fields store the gradient of the error * with respect to themselves * @param propagate_down * a vector with equal length to bottom, with each index indicating * whether to propagate the error gradients down to the bottom blob at * the corresponding index * @param bottom * the input blobs, whose diff fields will store the gradient of the error * with respect to themselves after Backward is run * * The Backward wrapper calls the relevant device wrapper function * (Backward_cpu or Backward_gpu) to compute the bottom blob diffs given the * top blob diffs. * * Your layer should implement Backward_cpu and (optionally) Backward_gpu. */  
      inline void Backward(const vector<Blob<Dtype>*>& top,  
          const vector<bool>& propagate_down,  
          const vector<Blob<Dtype>*>& bottom);  

     ...  

     protected:  
      /** The protobuf that stores the layer parameters */  
      LayerParameter layer_param_;  
      /** The phase: TRAIN or TEST */  
      Phase phase_;  
      /** The vector that stores the learnable parameters as a set of blobs. */  
      vector<shared_ptr<Blob<Dtype> > > blobs_;  
      /** Vector indicating whether to compute the diff of each param blob. */  
      vector<bool> param_propagate_down_;  

      /** The vector that indicates whether each top blob has a non-zero weight in * the objective function. */  
      vector<Dtype> loss_;  

      virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom,  
          const vector<Blob<Dtype>*>& top) = 0;  

      virtual void Forward_gpu(const vector<Blob<Dtype>*>& bottom,  
          const vector<Blob<Dtype>*>& top) {  
        // LOG(WARNING) << "Using CPU code as backup."; 
        return Forward_cpu(bottom, top);  
      }  

      virtual void Backward_cpu(const vector<Blob<Dtype>*>& top,  
          const vector<bool>& propagate_down,  
          const vector<Blob<Dtype>*>& bottom) = 0;  

      virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,  
          const vector<bool>& propagate_down,  
          const vector<Blob<Dtype>*>& bottom) {  
        // LOG(WARNING) << "Using CPU code as backup."; 
        Backward_cpu(top, propagate_down, bottom);  
      }  

    ...  

     };  // class Layer

说明：每一层定义了三种操作

Setup：Layer的初始化
Forward：前向传导计算，根据bottom计算top，调用了Forward_cpu（必须实现）和Forward_gpu（可选，若未实现，则调用cpu的）
Backward：反向传导计算，根据top计算bottom的梯度，其他同上

2.4 派生类分类

在Layer的派生类中，主要可以分为Vision Layers

Vision Layers
Vison 层主要用于处理视觉图像相关的层，以图像作为输入，产生其他的图像。其主要特点是具有空间结构。
包含Convolution(conv_layer.hpp)、Pooling(pooling_layer.hpp)、Local Response Normalization(LRN)(lrn_layer.hpp)、im2col等，注：老版本的Caffe有头文件include/caffe/vision_layers.hpp，新版本中用include/caffe/layer/conv_layer.hpp等取代
Loss Layers
这些层产生loss，如Softmax(SoftmaxWithLoss)、Sum-of-Squares / Euclidean(EuclideanLoss)、Hinge / Margin(HingeLoss)、Sigmoid Cross-Entropy(SigmoidCrossEntropyLoss)、Infogain(InfogainLoss)、Accuracy and Top-k等
Activation / Neuron Layers
元素级别的运算，运算均为同址计算（in-place computation，返回值覆盖原值而占用新的内存）。如：ReLU / Rectified-Linear and Leaky-ReLU(ReLU)、Sigmoid(Sigmoid)、TanH / Hyperbolic Tangent(TanH)、Absolute Value(AbsVal)、Power(Power)、BNLL(BNLL)等
Data Layers
网络的最底层，主要实现数据格式的转换，如：Database(Data)、In-Memory(MemoryData)、HDF5 Input(HDF5Data)、HDF5 Output(HDF5Output)、Images(ImageData)、Windows(WindowData)、Dummy(DummyData)等
Common Layers
Caffe提供了单个层与多个层的连接。如：Inner Product(InnerProduct)、Splitting(Split)、Flattening(Flatten)、Reshape(Reshape)、Concatenation(Concat)、Slicing(Slice)、Elementwise(Eltwise)、Argmax(ArgMax)、Softmax(Softmax)、Mean-Variance Normalization(MVN)等

注，括号内为Layer Type,没有括号暂缺信息，详细咱见引用2

3 Net

3.1 简介

一个Net由多个Layer组成。一个典型的网络从data layer（从磁盘中载入数据）出发到loss layer结束。如图是一个简单的逻辑回归分类器。

如下定义：

name: "LogReg"
layer { name: "mnist" type: "Data" top: "data" top: "label" data_param { source: "input_leveldb" batch_size: 64 }
}
layer { name: "ip" type: "InnerProduct" bottom: "data" top: "ip" inner_product_param { num_output: 2 }
}
layer { name: "loss" type: "SoftmaxWithLoss" bottom: "ip" bottom: "label" top: "loss" }

3.2 源代码

/** * @brief Connects Layer%s together into a directed acyclic graph (DAG) * specified by a NetParameter. * * TODO(dox): more thorough description. */
template <typename Dtype>
class Net {
 public:
...
  /// @brief Initialize a network with a NetParameter.
  void Init(const NetParameter& param);
...

  const vector<Blob<Dtype>*>& Forward(const vector<Blob<Dtype>* > & bottom,
      Dtype* loss = NULL);
...
  /** * The network backward should take no input and output, since it solely * computes the gradient w.r.t the parameters, and the data has already been * provided during the forward pass. */
  void Backward();
  ...
  Dtype ForwardBackward(const vector<Blob<Dtype>* > & bottom) {
    Dtype loss;
    Forward(bottom, &loss);
    Backward();
    return loss;
  }
...

 protected:
  ...
  /// @brief The network name
  string name_;
  /// @brief The phase: TRAIN or TEST
  Phase phase_;
  /// @brief Individual layers in the net
  vector<shared_ptr<Layer<Dtype> > > layers_;
  /// @brief the blobs storing intermediate results between the layer.
  vector<shared_ptr<Blob<Dtype> > > blobs_;
  vector<vector<Blob<Dtype>*> > bottom_vecs_;
  vector<vector<Blob<Dtype>*> > top_vecs_;
  ...
  /// The root net that actually holds the shared layers in data parallelism
  const Net* const root_net_;
};
}  // namespace caffe