caffe train 整体流程图

程序入口：main()

int main(int argc, char** argv) {
     .....
     return GetBrewFunction(caffe::string(argv[1]))();
     ....
}

g_brew_map实现过程，首先通过 typedef定义函数指针 typedef int (*BrewFunction)(); 这个是用typedef定义函数指针方法。这个程序定义一个BrewFunction函数指针类型，在caffe.cpp 中 BrewFunction 作为GetBrewFunction()函数的返回类型，可以是 train()，test()，device_query()，time() 这四个函数指针的其中一个。在train()，test()，中可以调用solver类的函数，从而进入到net，进入到每一层，运行整个caffe程序。然后对每个函数注册。

1. RegisterBrewFunction(train)
2. RegisterBrewFunction(test)
3. RegisterBrewFunction(device_query)
4. RegisterBrewFunction(time)
    train: 训练或者调整一个模型
    test : 在测试集上测试一个模型
    device_query : 打印GPU的调试信息
    time: 压测一个模型的执行时间

如果需要，可以增加其他的方式，然后通过RegisterBrewFunction()函数注册一下即可。

调用train()函数

接着调用train()函数，train函数中主要有三个方法ReadSolverParamsFromTextFileOrDie、CreateSolver、Solve。

// Train / Finetune a model.
int train() {
  ......
  caffe::SolverParameter solver_param;
  caffe::ReadSolverParamsFromTextFileOrDie(FLAGS_solver, &solver_param);//从-solver参数读取solver_param
  ......
  shared_ptr<caffe::Solver<float> >
      solver(caffe::SolverRegistry<float>::CreateSolver(solver_param));//从参数创建solver，同样采用string到函数指针的映射实现，用到了工厂模式

  if (FLAGS_snapshot.size()) {//迭代snapshot次后保存模型一次
    LOG(INFO) << "Resuming from " << FLAGS_snapshot;
    solver->Restore(FLAGS_snapshot.c_str());
  } else if (FLAGS_weights.size()) {//若采用finetuning，则拷贝weight到指定模型
    CopyLayers(solver.get(), FLAGS_weights);
  }

  if (gpus.size() > 1) {
    caffe::P2PSync<float> sync(solver, NULL, solver->param());
    sync.Run(gpus);
  } else {
    LOG(INFO) << "Starting Optimization";
    solver->Solve();//开始训练网络
  }
  LOG(INFO) << "Optimization Done.";
  return 0;
}

ReadSolverParamsFromTextFileOrDie

caffe::ReadSolverParamsFromTextFileOrDie(FLAGS_solver, &solver_param)解析-solver指定的solver.prototxt的文件内容到solver_param中

CreateSolver

CreateSolver函数构建solver和net，该函数是初始化的入口，会通过执行Solver的构造函数，调用 void Solver<Dtype>::Init(const SolverParameter& param)，该函数内有InitTrainNet()、InitTestNets()。对于InitTrainNet函数：

......
net_.reset(new Net<Dtype>(net_param));

调用Net类的构造函数，然后执行Init()操作，该函数具体的内容如下图和源码所示：

template <typename Dtype>
void Net<Dtype>::Init(const NetParameter& in_param) {
  ........//过滤校验参数FilterNet
  FilterNet(in_param, &filtered_param);
  .........//插入Splits层
  InsertSplits(filtered_param, &param);
  .......// 构建网络中输入输出存储结构
  bottom_vecs_.resize(param.layer_size());
  top_vecs_.resize(param.layer_size());
  bottom_id_vecs_.resize(param.layer_size());
  param_id_vecs_.resize(param.layer_size());
  top_id_vecs_.resize(param.layer_size());
  bottom_need_backward_.resize(param.layer_size());

  for (int layer_id = 0; layer_id < param.layer_size(); ++layer_id) {
   ...//创建层
 layers_.push_back(LayerRegistry<Dtype>::CreateLayer(layer_param));
    layer_names_.push_back(layer_param.name());
    LOG_IF(INFO, Caffe::root_solver())
        << "Creating Layer " << layer_param.name();
    bool need_backward = false;

    // 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);


   ........//创建相关blob
    // If the layer specifies that AutoTopBlobs() -> true and the LayerParameter
    // specified fewer than the required number (as specified by
    // ExactNumTopBlobs() or MinTopBlobs()), allocate them here.
    Layer<Dtype>* layer = layers_[layer_id].get();
    if (layer->AutoTopBlobs()) {
      const int needed_num_top =
          std::max(layer->MinTopBlobs(), layer->ExactNumTopBlobs());
      for (; num_top < needed_num_top; ++num_top) {
        // Add "anonymous" top blobs -- do not modify available_blobs or
        // blob_name_to_idx as we don't want these blobs to be usable as input
        // to other layers.
        AppendTop(param, layer_id, num_top, NULL, NULL);
      }
    }


    .....//执行SetUp()
    // After this layer is connected, set it up.
    layers_[layer_id]->SetUp(bottom_vecs_[layer_id], top_vecs_[layer_id]);
    LOG_IF(INFO, Caffe::root_solver())
        << "Setting up " << layer_names_[layer_id];
    for (int top_id = 0; top_id < top_vecs_[layer_id].size(); ++top_id) {
      if (blob_loss_weights_.size() <= top_id_vecs_[layer_id][top_id]) {
        blob_loss_weights_.resize(top_id_vecs_[layer_id][top_id] + 1, Dtype(0));
      }
      blob_loss_weights_[top_id_vecs_[layer_id][top_id]] = layer->loss(top_id);
      LOG_IF(INFO, Caffe::root_solver())
          << "Top shape: " << top_vecs_[layer_id][top_id]->shape_string();
      if (layer->loss(top_id)) {
        LOG_IF(INFO, Caffe::root_solver())
            << "    with loss weight " << layer->loss(top_id);
      }
      memory_used_ += top_vecs_[layer_id][top_id]->count();
    }
    LOG_IF(INFO, Caffe::root_solver())
        << "Memory required for data: " << memory_used_ * sizeof(Dtype);
    const int param_size = layer_param.param_size();
    const int num_param_blobs = layers_[layer_id]->blobs().size();
    CHECK_LE(param_size, num_param_blobs)
        << "Too many params specified for layer " <<

Net::Init()

SetUp是怎么构建的呢？

virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top) {}

 void SetUp(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top) {
    InitMutex();
    CheckBlobCounts(bottom, top);
    LayerSetUp(bottom, top);
    Reshape(bottom, top);
    SetLossWeights(top);
  }

初始化的总体流程大概就是新建一个Solver对象，然后调用Solver类的构造函数，然后在Solver的构造函数中又会新建Net类实例，在Net类的构造函数中又会新建各个layer的实例，一直具体到设置每个Blob，大概就完成了网络初始化的工作了。

Solve

train函数中CreateSolver()执行完成后，接下来是具体训练过程，执行Solve()函数---->Step()--->结束

Solve的具体内容和代码：

template <typename Dtype>
void Solver<Dtype>::Solve(const char* resume_file) {
  CHECK(Caffe::root_solver());
  LOG(INFO) << "Solving " << net_->name();
  LOG(INFO) << "Learning Rate Policy: " << param_.lr_policy();

  // For a network that is trained by the solver, no bottom or top vecs
  // should be given, and we will just provide dummy vecs.
  int start_iter = iter_;
  Step(param_.max_iter() - iter_);

  // overridden by setting snapshot_after_train := false
  if (param_.snapshot_after_train()
      && (!param_.snapshot() || iter_ % param_.snapshot() != 0)) {
    Snapshot();
  }

  // display loss
  if (param_.display() && iter_ % param_.display() == 0) {
    int average_loss = this->param_.average_loss();
    Dtype loss;
    net_->Forward(&loss);

    UpdateSmoothedLoss(loss, start_iter, average_loss);


  if (param_.test_interval() && iter_ % param_.test_interval() == 0) {
    TestAll();
  }
}

然后开始执行Step函数，具体内容和代码：

template <typename Dtype>  
void Solver<Dtype>::Step(int iters)  
{  
    // 起始迭代步数  
    const int start_iter = iter_;  
    // 终止迭代步数  
    const int stop_iter = iter_ + iters;  

    // 判断是否已经完成设定步数  
    while (iter_ < stop_iter)  
    {  
        // 将net_中的Bolb梯度参数置为零  
        net_->ClearParamDiffs();  

        ...  

        // accumulate the loss and gradient  
        Dtype loss = 0;  
        for (int i = 0; i < param_.iter_size(); ++i)  
        {  
            // 正向传导和反向传导，并计算loss  
            loss += net_->ForwardBackward();  
        }  
        loss /= param_.iter_size();  

        // 为了输出结果平滑，将临近的average_loss个loss数值进行平均，存储在成员变量smoothed_loss_中  
        UpdateSmoothedLoss(loss, start_iter, average_loss);  

        // BP算法更新权重  
        ApplyUpdate();  

        // Increment the internal iter_ counter -- its value should always indicate  
        // the number of times the weights have been updated.  
        ++iter_;  
    }  
}

while循环中先调用了网络类Net::ForwardBackward()成员函数进行正向传导和反向传导，并计算loss

Dtype ForwardBackward() {
    Dtype loss;
    //正向传导
    Forward(&loss);
    //反向传导
    Backward();
    return loss;
  }

而Fordward函数中调用了ForwardFromTo，而FordwardFromTo又调用了每个layer的Fordward。反向传导函数Backward()调用了BackwardFromTo(int start, int end)函数。正向传导和反向传导结束后，再调用SGDSolver::ApplyUpdate()成员函数进行权重更新。

• ForwardBackward：按顺序调用了Forward和Backward。
• ForwardFromTo(int start, int end)：执行从start层到end层的前向传递，采用简单的for循环调用。，forward只要计算损失loss
• BackwardFromTo(int start, int end)：和前面的ForwardFromTo函数类似，调用从start层到end层的反向传递。backward主要根据loss来计算梯度，caffe通过自动求导并反向组合每一层的梯度来计算整个网络的梯度。
• ToProto函数完成网络的序列化到文件，循环调用了每个层的ToProto函数

template <typename Dtype>  
void SGDSolver<Dtype>::ApplyUpdate()  
{  
    // 获取当前学习速率  
    Dtype rate = GetLearningRate();  
    if (this->param_.display() && this->iter_ % this->param_.display() == 0)  
    {  
        LOG(INFO) << "Iteration " << this->iter_ << ", lr = " << rate;  
    }  

    // 在计算当前梯度的时候，如果该值超过了阈值clip_gradients，则将梯度直接设置为该阈值  
    // 此处阈值设为-1，即不起作用  
    ClipGradients();  

    // 逐层更新网络中的可学习层  
    for (int param_id = 0; param_id < this->net_->learnable_params().size();  
       ++param_id)  
    {  
        // 归一化  
        Normalize(param_id);  
        // L2范数正则化添加衰减权重  
        Regularize(param_id);  
        // 随机梯度下降法计算更新值  
        ComputeUpdateValue(param_id, rate);  
    }  
    // 更新权重  
    this->net_->Update();  
}

ApplyUpdate

最后将迭代次数++iter_，继续while循环，直到迭代次数完成。 这就是整个网络的训练过程。