Caffe源码解析10：Caffe的求解（Solver）

1、先看一个Caffe Model Zoo的例子

  cd到当前Caffe目录，下载meta数据，通过data\ilsvrc12\get_ilsvrc_aux.sh脚本，windows用户直接通过脚本中“http://dl.caffe.berkeleyvision.org/caffe_ilsvrc12.tar.gz”下载该数据包，解压。 将其中的imagenet_mean.binaryproto和synset_words.txt文件放到data\ilsvrc12\目录下面。

  下载caffenet模型，通过http://dl.caffe.berkeleyvision.org/bvlc_reference_caffenet.caffemodel, 将文件放到“D:\VS2012\Projects\caffe-windows\models\bvlc_reference_caffenet”

  运行命令“D:\VS2012\Projects\caffe-windows>Build\x64\Release\classification.exe models\bvlc_reference_caffenet\deploy.prototxt models\bvlc_reference_caffenet\bvlc_reference_caffenet.caffemodel data\ilsvrc12\imagenet_mean.binaryproto data\ilsvrc12\synset_words.txt "examples\images\cat.jpg"“

得到

---------- Prediction for examples\images\cat.jpg ----------
0.3134 - "n02123045 tabby, tabby cat"
0.2380 - "n02123159 tiger cat"
0.1235 - "n02124075 Egyptian cat"
0.1003 - "n02119022 red fox, Vulpes vulpes"
0.0715 - "n02127052 lynx, catamount"

查看了classification.cpp的main（）可知该命令的运用需5个参数

int main(int argc, char** argv) {
  if (argc != 6) {
    std::cerr << "Usage: " << argv[0]
              << " deploy.prototxt network.caffemodel"
              << " mean.binaryproto labels.txt img.jpg" << std::endl;
    return 1;
  }

  ::google::InitGoogleLogging(argv[0]);

  string model_file   = argv[1];
  string trained_file = argv[2];
  string mean_file    = argv[3];
  string label_file   = argv[4];
  Classifier classifier(model_file, trained_file, mean_file, label_file);

  string file = argv[5];

  std::cout << "---------- Prediction for "
            << file << " ----------" << std::endl;

…

}

2、求解器是什么

求解器负责对模型优化，它的KPI就是让损失函数达到全局最小。

求解器特性如下：

• 记录优化过程，创建用于学习的训练网络和用于评估学习效果的测试网络
• 调用Forward->调用Backward->更新权值，反复迭代优化模型
• 周期性的评估测试网络
• 在优化过程中为模型、求解器状态打快照

为了让权值从初始态向跟好的模型前进，求解器在每次迭代中做以下事情：

• 调用Net的前向传播函数来计算输出和损失函数
• 调用Net的反向传播函数来计算梯度
• 根据求解器方法，将梯度转化为权值增量
• 根据学习速率、历史权值、所用方法更新求解器状态。

在设计求解器时，学习速率参数的设置主要针对卷积层和全连接层（统称权值层）

学习速率（learning rate）是负梯度的权重，遗忘因子（momentum）是权值更新历史的权重。顾名思义，学习速率和遗忘因子分别代表了今天学了多少新知识，以及以前学过的知识还记得多少。

3、求解器的实现

查看caffe.proto

message SolverParameter {
  //
  // Specifying the train and test networks
  //
  // Exactly one train net must be specified using one of the following fields:
  //     train_net_param, train_net, net_param, net
  // One or more test nets may be specified using any of the following fields:
  //     test_net_param, test_net, net_param, net
  // If more than one test net field is specified (e.g., both net and
  // test_net are specified), they will be evaluated in the field order given
  // above: (1) test_net_param, (2) test_net, (3) net_param/net.
  // A test_iter must be specified for each test_net.
  // A test_level and/or a test_stage may also be specified for each test_net.
  //

  // Proto filename for the train net, possibly combined with one or more
  // test nets.
  optional string net = 24;
  // Inline train net param, possibly combined with one or more test nets.
  optional NetParameter net_param = 25;

  optional string train_net = 1; // Proto filename for the train net.
  repeated string test_net = 2; // Proto filenames for the test nets.
  optional NetParameter train_net_param = 21; // Inline train net params.
  repeated NetParameter test_net_param = 22; // Inline test net params.

  // The states for the train/test nets. Must be unspecified or
  // specified once per net.
  //
  // By default, all states will have solver = true;
  // train_state will have phase = TRAIN,
  // and all test_state's will have phase = TEST.
  // Other defaults are set according to the NetState defaults.
  optional NetState train_state = 26;
  repeated NetState test_state = 27;

  // The number of iterations for each test net.
  repeated int32 test_iter = 3;

  // The number of iterations between two testing phases.
  optional int32 test_interval = 4 [default = 0];
  optional bool test_compute_loss = 19 [default = false];
  // If true, run an initial test pass before the first iteration,
  // ensuring memory availability and printing the starting value of the loss.
  optional bool test_initialization = 32 [default = true];
  optional float base_lr = 5; // The base learning rate
  // the number of iterations between displaying info. If display = 0, no info
  // will be displayed.
  optional int32 display = 6;
  // Display the loss averaged over the last average_loss iterations
  optional int32 average_loss = 33 [default = 1];
  optional int32 max_iter = 7; // the maximum number of iterations
  // accumulate gradients over `iter_size` x `batch_size` instances
  optional int32 iter_size = 36 [default = 1];

  // The learning rate decay policy. The currently implemented learning rate
  // policies are as follows:
  //    - fixed: always return base_lr.
  //    - step: return base_lr * gamma ^ (floor(iter / step))
  //    - exp: return base_lr * gamma ^ iter
  //    - inv: return base_lr * (1 + gamma * iter) ^ (- power)
  //    - multistep: similar to step but it allows non uniform steps defined by
  //      stepvalue
  //    - poly: the effective learning rate follows a polynomial decay, to be
  //      zero by the max_iter. return base_lr (1 - iter/max_iter) ^ (power)
  //    - sigmoid: the effective learning rate follows a sigmod decay
  //      return base_lr ( 1/(1 + exp(-gamma * (iter - stepsize))))
  //
  // where base_lr, max_iter, gamma, step, stepvalue and power are defined
  // in the solver parameter protocol buffer, and iter is the current iteration.
  optional string lr_policy = 8;
  optional float gamma = 9; // The parameter to compute the learning rate.
  optional float power = 10; // The parameter to compute the learning rate.
  optional float momentum = 11; // The momentum value.
  optional float weight_decay = 12; // The weight decay.
  // regularization types supported: L1 and L2
  // controlled by weight_decay
  optional string regularization_type = 29 [default = "L2"];
  // the stepsize for learning rate policy "step"
  optional int32 stepsize = 13;
  // the stepsize for learning rate policy "multistep"
  repeated int32 stepvalue = 34;

  // Set clip_gradients to >= 0 to clip parameter gradients to that L2 norm,
  // whenever their actual L2 norm is larger.
  optional float clip_gradients = 35 [default = -1];

  optional int32 snapshot = 14 [default = 0]; // The snapshot interval
  optional string snapshot_prefix = 15; // The prefix for the snapshot.
  // whether to snapshot diff in the results or not. Snapshotting diff will help
  // debugging but the final protocol buffer size will be much larger.
  optional bool snapshot_diff = 16 [default = false];
  enum SnapshotFormat {
    HDF5 = 0;
    BINARYPROTO = 1;
  }
  optional SnapshotFormat snapshot_format = 37 [default = BINARYPROTO];
  // the mode solver will use: 0 for CPU and 1 for GPU. Use GPU in default.
  enum SolverMode {
    CPU = 0;
    GPU = 1;
  }
  optional SolverMode solver_mode = 17 [default = GPU];
  // the device_id will that be used in GPU mode. Use device_id = 0 in default.
  optional int32 device_id = 18 [default = 0];
  // If non-negative, the seed with which the Solver will initialize the Caffe
  // random number generator -- useful for reproducible results. Otherwise,
  // (and by default) initialize using a seed derived from the system clock.
  optional int64 random_seed = 20 [default = -1];

  // type of the solver
  optional string type = 40 [default = "SGD"];

  // numerical stability for RMSProp, AdaGrad and AdaDelta and Adam
  optional float delta = 31 [default = 1e-8];
  // parameters for the Adam solver
  optional float momentum2 = 39 [default = 0.999];

  // RMSProp decay value
  // MeanSquare(t) = rms_decay*MeanSquare(t-1) + (1-rms_decay)*SquareGradient(t)
  optional float rms_decay = 38 [default = 0.99];

  // If true, print information about the state of the net that may help with
  // debugging learning problems.
  optional bool debug_info = 23 [default = false];

  // If false, don't save a snapshot after training finishes.
  optional bool snapshot_after_train = 28 [default = true];

  // DEPRECATED: old solver enum types, use string instead
  enum SolverType {
    SGD = 0;
    NESTEROV = 1;
    ADAGRAD = 2;
    RMSPROP = 3;
    ADADELTA = 4;
    ADAM = 5;
  }
  // DEPRECATED: use type instead of solver_type
  optional SolverType solver_type = 30 [default = SGD];
}

查看一个solver.prototxt

net: "models/bvlc_reference_caffenet/train_val.prototxt"
test_iter: 1000
test_interval: 1000
base_lr: 0.01        //基准学习速率为0.01，每个layer会在基准上进行细调
lr_policy: "step"    //学习速率衰减策略，step为步进方式，即每进行step次迭代，学习速率更新一次
gamma: 0.1           //学习速率衰减常数，每次更新学习速率都是乘上这个固定常数
stepsize: 100000     //每10万次迭代，对学习速率进行一次更新
display: 20          //每次打印信息的间隔，单位为迭代次数       
max_iter: 450000     //训练总共需要45万次迭代
momentum: 0.9        //遗忘因子
weight_decay: 0.0005  //权值衰减常数
snapshot: 10000       //打快照的间隔，单位为迭代次数
snapshot_prefix: "models/bvlc_reference_caffenet/caffenet_train"
solver_mode: GPU

查看include/caffe/solver.hpp

template <typename Dtype>
class Solver {
 public:
  explicit Solver(const SolverParameter& param,
      const Solver* root_solver = NULL);
  explicit Solver(const string& param_file, const Solver* root_solver = NULL);
  void Init(const SolverParameter& param);
  void InitTrainNet();
  void InitTestNets();

  // Client of the Solver optionally may call this in order to set the function
  // that the solver uses to see what action it should take (e.g. snapshot or
  // exit training early).
  void SetActionFunction(ActionCallback func);
  SolverAction::Enum GetRequestedAction();
  // The main entry of the solver function. In default, iter will be zero. Pass
  // in a non-zero iter number to resume training for a pre-trained net.
  virtual void Solve(const char* resume_file = NULL);
  inline void Solve(const string resume_file) { Solve(resume_file.c_str()); }
  void Step(int iters);
  // The Restore method simply dispatches to one of the
  // RestoreSolverStateFrom___ protected methods. You should implement these
  // methods to restore the state from the appropriate snapshot type.
  void Restore(const char* resume_file);
  // The Solver::Snapshot function implements the basic snapshotting utility
  // that stores the learned net. You should implement the SnapshotSolverState()
  // function that produces a SolverState protocol buffer that needs to be
  // written to disk together with the learned net.
  void Snapshot();
  virtual ~Solver() {}
  inline const SolverParameter& param() const { return param_; }
  inline shared_ptr<Net<Dtype> > net() { return net_; }
  inline const vector<shared_ptr<Net<Dtype> > >& test_nets() {
    return test_nets_;
  }
  int iter() { return iter_; }

  // Invoked at specific points during an iteration
  class Callback {
   protected:
    virtual void on_start() = 0;
    virtual void on_gradients_ready() = 0;

    template <typename T>
    friend class Solver;
  };
  const vector<Callback*>& callbacks() const { return callbacks_; }
  void add_callback(Callback* value) {
    callbacks_.push_back(value);
  }

  void CheckSnapshotWritePermissions();
  /**
   * @brief Returns the solver type.
   */
  virtual inline const char* type() const { return ""; }

 protected:
  // Make and apply the update value for the current iteration.
  virtual void ApplyUpdate() = 0;
  string SnapshotFilename(const string extension);
  string SnapshotToBinaryProto();
  string SnapshotToHDF5();
  // The test routine
  void TestAll();
  void Test(const int test_net_id = 0);
  virtual void SnapshotSolverState(const string& model_filename) = 0;
  virtual void RestoreSolverStateFromHDF5(const string& state_file) = 0;
  virtual void RestoreSolverStateFromBinaryProto(const string& state_file) = 0;
  void DisplayOutputBlobs(const int net_id);
  void UpdateSmoothedLoss(Dtype loss, int start_iter, int average_loss);

  SolverParameter param_;
  int iter_;
  int current_step_;
  shared_ptr<Net<Dtype> > net_;
  vector<shared_ptr<Net<Dtype> > > test_nets_;
  vector<Callback*> callbacks_;
  vector<Dtype> losses_;
  Dtype smoothed_loss_;

  // The root solver that holds root nets (actually containing shared layers)
  // in data parallelism
  const Solver* const root_solver_;

  // A function that can be set by a client of the Solver to provide indication
  // that it wants a snapshot saved and/or to exit early.
  ActionCallback action_request_function_;

  // True iff a request to stop early was received.
  bool requested_early_exit_;

  DISABLE_COPY_AND_ASSIGN(Solver);
};

用SGD方法的求解器，派生于Solver，查看sgd_solver.hpp, 其他类似的还有RMSPropSolver、AdaDeltaSolver、AdamSolver…等。

template <typename Dtype>
class SGDSolver : public Solver<Dtype> {
 public:
  explicit SGDSolver(const SolverParameter& param)
      : Solver<Dtype>(param) { PreSolve(); }
  explicit SGDSolver(const string& param_file)
      : Solver<Dtype>(param_file) { PreSolve(); }
  virtual inline const char* type() const { return "SGD"; }

  const vector<shared_ptr<Blob<Dtype> > >& history() { return history_; }

 protected:
  void PreSolve();
  Dtype GetLearningRate();
  virtual void ApplyUpdate();
  virtual void Normalize(int param_id);
  virtual void Regularize(int param_id);
  virtual void ComputeUpdateValue(int param_id, Dtype rate);
  virtual void ClipGradients();
  virtual void SnapshotSolverState(const string& model_filename);
  virtual void SnapshotSolverStateToBinaryProto(const string& model_filename);
  virtual void SnapshotSolverStateToHDF5(const string& model_filename);
  virtual void RestoreSolverStateFromHDF5(const string& state_file);
  virtual void RestoreSolverStateFromBinaryProto(const string& state_file);
  // history maintains the historical momentum data.
  // update maintains update related data and is not needed in snapshots.
  // temp maintains other information that might be needed in computation
  //   of gradients/updates and is not needed in snapshots
  vector<shared_ptr<Blob<Dtype> > > history_, update_, temp_;

  DISABLE_COPY_AND_ASSIGN(SGDSolver);
};

查看求解器的核心程序

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();

  // Initialize to false every time we start solving.
  requested_early_exit_ = false;

  if (resume_file) {  //如果指定了快照恢复文件，则从快照恢复训练环境
    LOG(INFO) << "Restoring previous solver status from " << resume_file;
    Restore(resume_file);
  }

  // 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_);  //关键函数
  // If we haven't already, save a snapshot after optimization, unless
  // overridden by setting snapshot_after_train := false
  if (param_.snapshot_after_train()
      && (!param_.snapshot() || iter_ % param_.snapshot() != 0)) {
    Snapshot();
  }
  if (requested_early_exit_) {
    LOG(INFO) << "Optimization stopped early.";
    return;
  }
  // After the optimization is done, run an additional train and test pass to
  // display the train and test loss/outputs if appropriate (based on the
  // display and test_interval settings, respectively).  Unlike in the rest of
  // training, for the train net we only run a forward pass as we've already
  // updated the parameters "max_iter" times -- this final pass is only done to
  // display the loss, which is computed in the forward pass.
  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);

    LOG(INFO) << "Iteration " << iter_ << ", loss = " << smoothed_loss_;
  }
  if (param_.test_interval() && iter_ % param_.test_interval() == 0) {
    TestAll();
  }
  LOG(INFO) << "Optimization Done.";
}

深入Step()函数

template <typename Dtype>
void Solver<Dtype>::Step(int iters) {
  const int start_iter = iter_;
  const int stop_iter = iter_ + iters;
  int average_loss = this->param_.average_loss();
  losses_.clear();
  smoothed_loss_ = 0;

  while (iter_ < stop_iter) {
    // zero-init the params
    net_->ClearParamDiffs();
    if (param_.test_interval() && iter_ % param_.test_interval() == 0
        && (iter_ > 0 || param_.test_initialization())
        && Caffe::root_solver()) {
      TestAll();
      if (requested_early_exit_) {
        // Break out of the while loop because stop was requested while testing.
        break;
      }
    }

    for (int i = 0; i < callbacks_.size(); ++i) {
      callbacks_[i]->on_start();
    }
    const bool display = param_.display() && iter_ % param_.display() == 0;
    net_->set_debug_info(display && param_.debug_info());
    // accumulate the loss and gradient
    Dtype loss = 0;
    for (int i = 0; i < param_.iter_size(); ++i) {
      loss += net_->ForwardBackward();
    }
    loss /= param_.iter_size();
    // average the loss across iterations for smoothed reporting
    UpdateSmoothedLoss(loss, start_iter, average_loss);
    if (display) {
      LOG_IF(INFO, Caffe::root_solver()) << "Iteration " << iter_
          << ", loss = " << smoothed_loss_;
      const vector<Blob<Dtype>*>& result = net_->output_blobs();
      int score_index = 0;
      for (int j = 0; j < result.size(); ++j) {
        const Dtype* result_vec = result[j]->cpu_data();
        const string& output_name =
            net_->blob_names()[net_->output_blob_indices()[j]];
        const Dtype loss_weight =
            net_->blob_loss_weights()[net_->output_blob_indices()[j]];
        for (int k = 0; k < result[j]->count(); ++k) {
          ostringstream loss_msg_stream;
          if (loss_weight) {
            loss_msg_stream << " (* " << loss_weight
                            << " = " << loss_weight * result_vec[k] << " loss)";
          }
          LOG_IF(INFO, Caffe::root_solver()) << "    Train net output #"
              << score_index++ << ": " << output_name << " = "
              << result_vec[k] << loss_msg_stream.str();
        }
      }
    }
    for (int i = 0; i < callbacks_.size(); ++i) {
      callbacks_[i]->on_gradients_ready();
    }
    ApplyUpdate();

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

    SolverAction::Enum request = GetRequestedAction();

    // Save a snapshot if needed.
    if ((param_.snapshot()
         && iter_ % param_.snapshot() == 0
         && Caffe::root_solver()) ||
         (request == SolverAction::SNAPSHOT)) {
      Snapshot();
    }
    if (SolverAction::STOP == request) {
      requested_early_exit_ = true;
      // Break out of training loop.
      break;
    }
  }
}

由于ApplyUpdate()是纯虚函数，所以查看SGDSolver中的ApplyUpdate()

template <typename Dtype>
void SGDSolver<Dtype>::ApplyUpdate() {
  CHECK(Caffe::root_solver());
  Dtype rate = GetLearningRate();
  if (this->param_.display() && this->iter_ % this->param_.display() == 0) {
    LOG(INFO) << "Iteration " << this->iter_ << ", lr = " << rate;
  }
  ClipGradients(); //钳制误差梯度
  for (int param_id = 0; param_id < this->net_->learnable_params().size();  //对训练网络中的每个权值Blob，都做归一化、正则化、计算增量这三步
       ++param_id) {
    Normalize(param_id);
    Regularize(param_id);
    ComputeUpdateValue(param_id, rate);
  }
  this->net_->Update();
}

Solver每隔一定周期会对训练的网络做一次评估，使用test_nets_切换到新数据集进行预测，调用TestAll()函数：

template <typename Dtype>
void Solver<Dtype>::TestAll() {
  for (int test_net_id = 0;
       test_net_id < test_nets_.size() && !requested_early_exit_;
       ++test_net_id) {
    Test(test_net_id);
  }
}

对单个网络的预测会用Test()函数:

template <typename Dtype>
void Solver<Dtype>::Test(const int test_net_id) {
  CHECK(Caffe::root_solver());
  LOG(INFO) << "Iteration " << iter_
            << ", Testing net (#" << test_net_id << ")";
  CHECK_NOTNULL(test_nets_[test_net_id].get())->
      ShareTrainedLayersWith(net_.get());
  vector<Dtype> test_score;
  vector<int> test_score_output_id;
  const shared_ptr<Net<Dtype> >& test_net = test_nets_[test_net_id];
  Dtype loss = 0;
  for (int i = 0; i < param_.test_iter(test_net_id); ++i) {
    SolverAction::Enum request = GetRequestedAction();
    // Check to see if stoppage of testing/training has been requested.
    while (request != SolverAction::NONE) {
        if (SolverAction::SNAPSHOT == request) {
          Snapshot();
        } else if (SolverAction::STOP == request) {
          requested_early_exit_ = true;
        }
        request = GetRequestedAction();
    }
    if (requested_early_exit_) {
      // break out of test loop.
      break;
    }

    Dtype iter_loss;
    const vector<Blob<Dtype>*>& result =
        test_net->Forward(&iter_loss);
    if (param_.test_compute_loss()) {
      loss += iter_loss;
    }
    if (i == 0) {
      for (int j = 0; j < result.size(); ++j) {
        const Dtype* result_vec = result[j]->cpu_data();
        for (int k = 0; k < result[j]->count(); ++k) {
          test_score.push_back(result_vec[k]);
          test_score_output_id.push_back(j);
        }
      }
    } else {
      int idx = 0;
      for (int j = 0; j < result.size(); ++j) {
        const Dtype* result_vec = result[j]->cpu_data();
        for (int k = 0; k < result[j]->count(); ++k) {
          test_score[idx++] += result_vec[k];
        }
      }
    }
  }
  if (requested_early_exit_) {
    LOG(INFO)     << "Test interrupted.";
    return;
  }
  if (param_.test_compute_loss()) {
    loss /= param_.test_iter(test_net_id);
    LOG(INFO) << "Test loss: " << loss;
  }
  for (int i = 0; i < test_score.size(); ++i) {
    const int output_blob_index =
        test_net->output_blob_indices()[test_score_output_id[i]];
    const string& output_name = test_net->blob_names()[output_blob_index];
    const Dtype loss_weight = test_net->blob_loss_weights()[output_blob_index];
    ostringstream loss_msg_stream;
    const Dtype mean_score = test_score[i] / param_.test_iter(test_net_id);
    if (loss_weight) {
      loss_msg_stream << " (* " << loss_weight
                      << " = " << loss_weight * mean_score << " loss)";
    }
    LOG(INFO) << "    Test net output #" << i << ": " << output_name << " = "
              << mean_score << loss_msg_stream.str();
  }
}