本文從CSDN上轉(zhuǎn)移過來地址：
http://blog.csdn.net/mounty_fsc/article/details/51090306

1. Solver到Net

在SGDSolver的構(gòu)造函數(shù)中詳見本系列博文（二）陪白，主要執(zhí)行了其父類Solver的構(gòu)造函數(shù)夷野，接著執(zhí)行Solver::Init()函數(shù)入客，在Init()中，有兩個(gè)函數(shù)值得注意：InitTrainNet()和InitTestNets()分別初始化訓(xùn)練網(wǎng)絡(luò)和測試網(wǎng)絡(luò)背桐。

• InitTrainNet
  • 首先，ReadNetParamsFromTextFileOrDie(param_.net(), &net_param)把param_.net()（即examples/mnist/lenet_train_test.prototxt）中的信息讀入net_param蝉揍。
  • 其次链峭，net_.reset(new Net<Dtype>(net_param))重新構(gòu)建網(wǎng)絡(luò)，調(diào)用Net的構(gòu)造方法又沾。
  • 然后弊仪，在構(gòu)造方法中執(zhí)行Net::init()，開始正式創(chuàng)建網(wǎng)絡(luò)杖刷。其主要代碼如下：

template <typename Dtype>
void Net<Dtype>::Init(const NetParameter& in_param) {
...
for (int layer_id = 0; layer_id < param.layer_size(); ++layer_id) {

    // Setup layer.
    const LayerParameter& layer_param = param.layer(layer_id);
 
    layers_.push_back(LayerRegistry<Dtype>::CreateLayer(layer_param));

    // Figure out this layer's input and output
    for (int bottom_id = 0; bottom_id < layer_param.bottom_size();  ++bottom_id) {
      const int blob_id = AppendBottom(param, layer_id, bottom_id, &available_blobs, &blob_name_to_idx);
      // If a blob needs backward, this layer should provide it.
      need_backward |= blob_need_backward_[blob_id];
    }
    int num_top = layer_param.top_size();
    for (int top_id = 0; top_id < num_top; ++top_id) {
      AppendTop(param, layer_id, top_id, &available_blobs, &blob_name_to_idx);
    }
 ...

  layers_[layer_id]->SetUp(bottom_vecs_[layer_id], top_vecs_[layer_id]);
  ...
 }

for (int param_id = 0; param_id < num_param_blobs; ++param_id) {
  AppendParam(param, layer_id, param_id);
}

...
}
```

**說明：**

  1. Lenet5在caffe中共有9層励饵，即`param.layer_size() == 5`，以上代碼每一次for循環(huán)創(chuàng)建一個(gè)網(wǎng)絡(luò)層
  2. 每層網(wǎng)絡(luò)是通過`LayerRegistry::CreateLayer()`創(chuàng)建的挺勿，類似與Solver的創(chuàng)建（詳見本系列博文（二））
  3. 14行`Net::AppendBottom()`曲横，對(duì)于`layer_id`這層，從`Net::blob_`中取出blob放入該層對(duì)應(yīng)的`bottom_vecs_[layer_id]`中
  4. 20行`Net::AppendTop()`不瓶，對(duì)于`layer_id`這層禾嫉，創(chuàng)建`blob`（未包含數(shù)據(jù)）并放入`Net::blob_`中
  5. `Layer::SetUp()`
  6. `AppendParam`中把每層網(wǎng)絡(luò)的訓(xùn)練參數(shù)與網(wǎng)絡(luò)變量`learnable_params_`綁定，在lenet中蚊丐，只有`conv1`,`conv2`,`ip1`,`ip2`四層有參數(shù)熙参，每層分別有參數(shù)與偏置參數(shù)兩項(xiàng)參數(shù)，因而`learnable_params_`的size為8.

• InitTestNets
  <font color="red">該部分內(nèi)容見本系列博文：（Caffe麦备，Lenet5）初始化測試網(wǎng)絡(luò)（四）孽椰。</font>

2 訓(xùn)練網(wǎng)絡(luò)結(jié)構(gòu)

注：Top Blob Shape格式為：BatchSize，ChannelSize凛篙，Height黍匾，Width（Total Count）

3 第一層：Data Layer

3.1 protobuff定義

訓(xùn)練網(wǎng)絡(luò)的第一層protobuff定義為：

layer {
  name: "mnist"
  type: "Data"
  top: "data"
  top: "label"
  include {
    phase: TRAIN
  }
  transform_param {
    scale: 0.00390625
  }
  data_param {
    source: "examples/mnist/mnist_train_lmdb"
    batch_size: 64
    backend: LMDB
  }
}

3.2 函數(shù)LayerRegistry::CreateLayer

第1節(jié)中代碼第一次通過調(diào)用LayerRegistry::CreateLayer()創(chuàng)建了DataLayer類，DataLayer類的繼承關(guān)系如下圖所示呛梆，詳見[1]：

由繼承圖可知锐涯，調(diào)用DataLayer()的構(gòu)造函數(shù)，依次執(zhí)行的順序?yàn)槠浠悩?gòu)造函數(shù)：Layer()填物、BaseDataLayer()纹腌、InternalThread()（詳見（Caffe）基本類InternalThread（三） ）、BasePrefetchingDataLayer()滞磺、及DataLayer()升薯。

其中，值得注意的是DataLayer()击困，在調(diào)用基類構(gòu)造函數(shù)BasePrefetchingDataLayer()之后涎劈，對(duì) DataReader reader_ 進(jìn)行賦值，在該DataLayer對(duì)象中維護(hù)了一個(gè)DataReader對(duì)象reader_，其作用是添加讀取數(shù)據(jù)任務(wù)至炮障，一個(gè)專門讀取數(shù)據(jù)庫（examples/mnist/mnist_train_lmdb）的線程（若還不存在該線程目派，則創(chuàng)建該線程）坤候，此處一共取出了4*64個(gè)樣本至BlockingQueue<Datum*> DataReader::QueuePair::full_。詳見（Caffe）基本類DataReader企蹭、QueuePair白筹、Body（四）

template <typename Dtype>
DataLayer<Dtype>::DataLayer(const LayerParameter& param)
  : BasePrefetchingDataLayer<Dtype>(param),
    reader_(param) {
}

3.3 函數(shù)Layer::SetUp

• 此處按程序執(zhí)行順序值得關(guān)注的有：
  在DataLayer::DataLayerSetUp中根據(jù)3.2DataReader中介紹的讀取的數(shù)據(jù)中取出一個(gè)樣本推測blob的形狀

• BasePrefetchingDataLayer::LayerSetUp如下代碼prefetch_[i].data_.mutable_cpu_data()用到了涉及到gpu、cpu間復(fù)制數(shù)據(jù)的問題谅摄，見（Caffe）基本類Blob徒河，Layer，Net（一）1.4SyncedMemory及引用[2]

   // Before starting the prefetch thread, we make cpu_data and gpu_data
   // calls so that the prefetch thread does not accidentally make simultaneous
   // cudaMalloc calls when the main thread is running. In some GPUs this
   // seems to cause failures if we do not so.
   for (int i = 0; i < PREFETCH_COUNT; ++i) {
     prefetch_[i].data_.mutable_cpu_data();
     if (this->output_labels_) {
       prefetch_[i].label_.mutable_cpu_data();
     }
   }
  

• BasePrefetchingDataLayer類繼承了InternalThread送漠，BasePrefetchingDataLayer<Dtype>::LayerSetUp中通過調(diào)用StartInternalThread()開啟了一個(gè)新線程顽照，從而執(zhí)行BasePrefetchingDataLayer::InternalThreadEntry

• BasePrefetchingDataLayer::InternalThreadEntry關(guān)鍵代碼如下，其中load_batch(batch)為闽寡，從2.2介紹的BlockingQueue<Datum*> DataReader::QueuePair::full_（包含從數(shù)據(jù)庫讀出的數(shù)據(jù)）中讀取一個(gè)batch_size的數(shù)據(jù)到BlockingQueue<Batch<Dtype>*> BasePrefetchingDataLayer::prefetch_full_中代兵。由于該線程在prefetch_free_為空時(shí)將掛起等待（PREFETCH_COUNT=3），prefetch_full_中用完的Batch將放回prefetch_free_中爷狈。<u>該線程何時(shí)停止植影？</u>

      while (!must_stop()) {
        Batch<Dtype>* batch = prefetch_free_.pop();
        load_batch(batch);
  #ifndef CPU_ONLY
        if (Caffe::mode() == Caffe::GPU) {
          batch->data_.data().get()->async_gpu_push(stream);
          CUDA_CHECK(cudaStreamSynchronize(stream));
        }
  #endif
        prefetch_full_.push(batch);
      }
  

關(guān)于線程的總結(jié)：

1. 此外一共涉及到兩個(gè)線程，分別為都是繼承了InnerThread的BasePrefetchingDataLayer(DataLayer)類和DataReader中的Body類
2. Body為面向數(shù)據(jù)庫的線程涎永，不斷從某個(gè)數(shù)據(jù)庫中讀出數(shù)據(jù)思币，存放至緩存為隊(duì)列DataReader::QueuePair::BlockingQueue<Datum*>，一般保存4*64個(gè)單位數(shù)據(jù)羡微，單位為Datum
3. BasePrefetchingDataLayer為面向網(wǎng)絡(luò)的線程谷饿，從Body的緩存中不斷讀取數(shù)據(jù)。BasePrefetchingDataLayer的緩存為隊(duì)列BlockingQueue<Batch*>妈倔，一般存放3個(gè)單位的數(shù)據(jù)博投，單位為Batch

static const int PREFETCH_COUNT = 3;
Batch<Dtype> prefetch_[PREFETCH_COUNT];
BlockingQueue<Batch<Dtype>*> prefetch_free_;
BlockingQueue<Batch<Dtype>*> prefetch_full_;

template <typename Dtype>
BasePrefetchingDataLayer<Dtype>::BasePrefetchingDataLayer(
    const LayerParameter& param)
    : BaseDataLayer<Dtype>(param),
      prefetch_free_(), prefetch_full_() {
  for (int i = 0; i < PREFETCH_COUNT; ++i) {
    prefetch_free_.push(&prefetch_[i]);
  }
}

• prefetch_full_與prefetch_free_中的元素由prefetch_提供

4 第二層：Convolution Layer

4.1 protobuff定義

layer {
  name: "conv1"
  type: "Convolution"
  bottom: "data"
  top: "conv1"
  param {
    lr_mult: 1
  }
  param {
    lr_mult: 2
  }
  convolution_param {
    num_output: 20
    kernel_size: 5
    stride: 1
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
    }
  }
}

4.2 函數(shù)LayerRegistry::CreateLayer

這里寫圖片描述

說明：

1. 不像DataLayer 直接執(zhí)行的是構(gòu)造函數(shù)，此時(shí)執(zhí)行的是GetConvolutuionLayer()启涯，然后調(diào)用ConvolutionLayer()贬堵，原因如下：

REGISTER_LAYER_CREATOR(Convolution, GetConvolutionLayer);

4.3 Layer::SetUp

在Layer::SetUp中，調(diào)用了ConvolutionLayer的基類BaseConvolutionLayer的LayerSetUp及Reshape函數(shù)结洼，該類的主要成員變量如下：

/**
 * @brief Abstract base class that factors out the BLAS code common to
 *        ConvolutionLayer and DeconvolutionLayer.
 */
template <typename Dtype>
class BaseConvolutionLayer : public Layer<Dtype> {
 public:
  explicit BaseConvolutionLayer(const LayerParameter& param)
      : Layer<Dtype>(param) {}
  virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);
  virtual void Reshape(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);

 ...
  /// @brief The spatial dimensions of a filter kernel.
  Blob<int> kernel_shape_;
  /// @brief The spatial dimensions of the stride.
  Blob<int> stride_;
  /// @brief The spatial dimensions of the padding.
  Blob<int> pad_;
  /// @brief The spatial dimensions of the dilation.
  Blob<int> dilation_;
  /// @brief The spatial dimensions of the convolution input.
  Blob<int> conv_input_shape_;
  /// @brief The spatial dimensions of the col_buffer.
  vector<int> col_buffer_shape_;
  /// @brief The spatial dimensions of the output.
  vector<int> output_shape_;
  const vector<int>* bottom_shape_;
...
};

說明：

1. LayerSetUp函數(shù)中黎做，主要是初始化了kernel_shape_、stride_松忍、pad_蒸殿、dilation_以及初始化網(wǎng)絡(luò)參數(shù)，并存放與Layer::blobs_中。
2. Reshape函數(shù)中宏所，conv_input_shape_酥艳、bottom_shape_等

5 第三層：Pooling Layer

5.1 protobuff定義

layer {
  name: "pool1"
  type: "Pooling"
  bottom: "conv1"
  top: "pool1"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}

5.2 Layer::SetUp

通過調(diào)用虛函數(shù)LayerSetUp及Reshape對(duì)以下成員變量進(jìn)行初始化

/**
 * @brief Pools the input image by taking the max, average, etc. within regions.
 *
 * TODO(dox): thorough documentation for Forward, Backward, and proto params.
 */
template <typename Dtype>
class PoolingLayer : public Layer<Dtype> {
 ....
  int kernel_h_, kernel_w_;
  int stride_h_, stride_w_;
  int pad_h_, pad_w_;
  int channels_;
  int height_, width_;
  int pooled_height_, pooled_width_;
  bool global_pooling_;
  Blob<Dtype> rand_idx_;
  Blob<int> max_idx_;
};

6 第四層、第五層

基本同第二層爬骤、第三層

7 第六層：InnerProduct Layer

7.1 protobuff定義

layer {
  name: "ip1"
  type: "InnerProduct"
  bottom: "pool2"
  top: "ip1"
  param {
    lr_mult: 1
  }
  param {
    lr_mult: 2
  }
  inner_product_param {
    num_output: 500
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
    }
  }
}

7.2 Layer::SetUp

/**
 * @brief Also known as a "fully-connected" layer, computes an inner product
 *        with a set of learned weights, and (optionally) adds biases.
 *
 * TODO(dox): thorough documentation for Forward, Backward, and proto params.
 */
template <typename Dtype>
class InnerProductLayer : public Layer<Dtype> {
 ...
  int M_;
  int K_;
  int N_;
  bool bias_term_;
  Blob<Dtype> bias_multiplier_;
};

說明：

1. N_為輸出大小充石，即等于protobuff中定義的num_output
2. K_為輸入大小，對(duì)于該層Bottom Blob形狀為(N, C, H, W)霞玄，N為batch_size骤铃，K_=CHW（Caffe）基本類Blob，Layer坷剧，Net（一）惰爬，M_=N。其中只有C惫企、H撕瞧、W跟內(nèi)積相關(guān)

8 第七層：ReLU Layer

8.1 protobuff定義

layer {
  name: "relu1"
  type: "ReLU"
  bottom: "ip1"
  top: "ip1"
}

8.2 說明

ReLULayer主要是用來做計(jì)算的，其繼承關(guān)系如下狞尔，詳細(xì)參加[4]丛版、[5]

9 第八層：InnerProduct Layer

參見第7節(jié)

10 第九層：SoftmaxWithLoss Layer

10.1 protobuff定義

layer {
  name: "loss"
  type: "SoftmaxWithLoss"
  bottom: "ip2"
  bottom: "label"
  top: "loss"
}

10.2 LayerRegistry::CreateLayer

這里寫圖片描述

10.3 Layer::SetUp

值得注意的是：

1. 類SoftmaxWithLossLayer包含類SoftmaxLayer的實(shí)例
   shared_ptr<Layer<Dtype> > softmax_layer_

2. softmax_layer_在LayerSetUp中賦值。

3. 此函數(shù)內(nèi)調(diào)用Layer::SetLossWeights初始化了該層的Top Blob（loss）

4. 兩個(gè)類間的關(guān)系如下圖：

   這里寫圖片描述

5. 成員變量prob_作為Softmaxlayer的top blob

6. bottom blob[0]作為softmaxlayer的bottom blob

7. 所以經(jīng)過softmaxlayer計(jì)算之后沪么，得出64*10（每個(gè)樣本的每個(gè)類別上的概率）存放在prob_中

11 剩余的工作

至此硼婿，訓(xùn)練網(wǎng)絡(luò)基本創(chuàng)建完畢，接下來剩下的工作主要有：

1. 反向檢查一次網(wǎng)絡(luò)禽车，看哪些blobs會(huì)對(duì)loss產(chǎn)生影響寇漫，在LeNet5中，前面的9層均有影響
2. 初始化權(quán)值共享

[1].http://caffe.berkeleyvision.org/doxygen/classcaffe_1_1BasePrefetchingDataLayer.html
[2].http://caffe.berkeleyvision.org/tutorial/net_layer_blob.html Implementation Details
[3].http://caffe.berkeleyvision.org/doxygen/classcaffe_1_1ConvolutionLayer.html
[4].http://caffe.berkeleyvision.org/doxygen/classcaffe_1_1ReLULayer.html
[5].http://caffe.berkeleyvision.org/tutorial/layers.html ReLU / Rectified-Linear and Leaky-ReLU