caffe运行中显示梯度信息

为了显示各层的梯度变化，阅读了此篇博客http://blog.csdn.net/yihaizhiyan/article/details/44159063

但是新版caffe solver.cpp中做了优化，所以函数添加位置有变，在此记录下

修改目录 caffe/src/caffe/solver.cpp

开头修改如下：

#include <cstdio>


#include <algorithm>
#include <string>
#include <vector>
#include <iomanip>


#include "caffe/net.hpp"
#include "caffe/proto/caffe.pb.h"
#include "caffe/solver.hpp"
#include "caffe/util/io.hpp"
#include "caffe/util/math_functions.hpp"
#include "caffe/util/upgrade_proto.hpp"


using std::max;
using std::min;
using std::setw;
using std::scientific;
using std::left;

后面修改

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;
  iteration_timer_.Start();

  while (iter_ < stop_iter) {
    // zero-init the params
    net_->ClearParamDiffs();
    if (param_.test_interval() && iter_ % param_.test_interval() == 0
        && (iter_ > 0 || param_.test_initialization())) {
      if (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) {
      float lapse = iteration_timer_.Seconds();
      float per_s = (iter_ - iterations_last_) / (lapse ? lapse : 1);
      LOG_IF(INFO, Caffe::root_solver()) << "Iteration " << iter_
          << " (" << per_s << " iter/s, " << lapse << "s/"
          << param_.display() << " iters), loss = " << smoothed_loss_;
      iteration_timer_.Start();
      iterations_last_ = iter_;
      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();
        }
      }

       //add by me
      int max_len = 0;
      for (int l = 0; l < net_->layers().size(); ++l) {
        Layer<Dtype>& layer = *net_->layers()[l].get();
        if (layer.blobs().size() > 0 && layer.layer_param().name().length() > max_len)
         max_len = layer.layer_param().name().length();
         }
     for (int l = 0; l < net_->layers().size(); ++l) {
        Layer<Dtype>& layer = *net_->layers()[l].get();
        for (int b = 0; b < layer.blobs().size(); ++b) {
         Blob<Dtype>& blob = *layer.blobs()[b].get();
         const Dtype* blob_cpu_data = blob.cpu_data();
         const Dtype* blob_cpu_diff = blob.cpu_diff();
         Dtype data_sum = 0;
         Dtype diff_sum = 0;
         for (int i = 0; i < blob.count(); ++i) {
          data_sum += (blob_cpu_data[i] > Dtype(0.)) ? blob_cpu_data[i]
          : - blob_cpu_data[i];
          diff_sum += (blob_cpu_diff[i] > Dtype(0.)) ? blob_cpu_diff[i]
          : - blob_cpu_diff[i];
         }
         data_sum /= blob.count();
         diff_sum /= blob.count();
         LOG(INFO) << std::left << std::setw(max_len + 1) << std::setfill(' ')
                     << layer.layer_param().name()
                     << " blob" << b << ": " << std::scientific
            << data_sum << " [" << diff_sum << "]";
     }
    }
   //end revised by me
     
     
     
    }