Windows：Caffe 网络层及函数源码修改

一、Caffe pooling layer修改（添加ceil mode属性）

Step1：pooling_layer.hpp文件

1. 找到属性定义并添加bool ceil_mode_;

Step2：pooling_layer.cpp文件

1. 找到相应参数解析处添加代码

ceil_mode_ = param.ceil_mode()

2. 找到相应计算处根据个人需求添加代码

if ceil_mode_ ...
	else ...

Step3：caffe.proto文件

1. 找到pooling layer相应属性列表，添加

optional bool ceil_mode = 13 [default=true]

二、Caffe自定义图像处理

Step1：data_transformer.cpp修改

1. …\caffe-master\src\caffe中data_transformer.cpp文件找到如下两个函数并修改
2. 此函数主要是使用caffe Datalayer时调用的预处理函数

void DataTransformer<Dtype>::Transform(const Datum& datum, Dtype* transformed_data)

/// … ///之间为新写的代码，下方是被注释的原有代码

 
 // -------------------------------------新代码--------------------------------------------------
  Dtype mean_datum_element;
  int mean_top_index, mean_data_index;
  float pixel_sum_value = 0.0, global_mean = 0.0;

  for (int c = 0; c < datum_channels; ++c)
  {
	  for (int h = 0; h < height; ++h)
	  {
		  for (int w = 0; w < width; ++w)
		  {
			  mean_data_index = (c * datum_height + h_off + h) * datum_width + w_off + w;
			  mean_top_index = (c * height + h) * width + w;
			  if (has_uint8) {
				  mean_datum_element =
					  static_cast<Dtype>(static_cast<uint8_t>(data[mean_data_index]));
			  }
			  else {
				  mean_datum_element = datum.float_data(mean_data_index);
			  }
			  pixel_sum_value += mean_datum_element;

		  }
	  }
  }
  global_mean = pixel_sum_value / (datum_channels * height * width) * 1.0;
  //LOG(INFO) << "Using datum to load data ----------------------------------------------------";
  Dtype datum_element;
  int top_index, data_index;
  // 遍历像素点
  for (int c = 0; c < datum_channels; ++c) 
  {
    for (int h = 0; h < height; ++h) 
	{
      for (int w = 0; w < width; ++w) 
	  {
        data_index = (c * datum_height + h_off + h) * datum_width + w_off + w;
        top_index = (c * height + h) * width + w;
        if (has_uint8) {
          datum_element =
            static_cast<Dtype>(static_cast<uint8_t>(data[data_index]));
        } else {
          datum_element = datum.float_data(data_index);
        }
        transformed_data[top_index] = (datum_element * 1.0 - global_mean) * scale;
      }
	}
  }
//
  // ----------------------------------------------原有代码---------------------------------
  //Dtype datum_element;
  //int top_index, data_index;
  //for (int c = 0; c < datum_channels; ++c) {
  //  for (int h = 0; h < height; ++h) {
  //    for (int w = 0; w < width; ++w) {
  //      data_index = (c * datum_height + h_off + h) * datum_width + w_off + w;
  //      if (do_mirror) {
  //        top_index = (c * height + h) * width + (width - 1 - w);
  //      } else {
  //        top_index = (c * height + h) * width + w;
  //      }
  //      if (has_uint8) {
  //        datum_element =
  //          static_cast<Dtype>(static_cast<uint8_t>(data[data_index]));
  //      } else {
  //        datum_element = datum.float_data(data_index);
  //      }
  //      if (has_mean_file) {
  //        transformed_data[top_index] =
  //          (datum_element - mean[data_index]) * scale;
  //      } else {
  //        if (has_mean_values) {
  //          transformed_data[top_index] =
  //            (datum_element - mean_values_[c]) * scale;
  //        } else {
  //          transformed_data[top_index] = datum_element * scale;
  //        }
  //      }
  //    }
  //  }
  //}

3. 此函数主要是使用caffe imagedatalayer时调用的预处理函数

void DataTransformer<Dtype>::Transform(const cv::Mat& cv_img,  Blob<Dtype>* transformed_blob) 

 //
  // ----------------------------------------------新代码---------------------------------
  cv::Scalar channel_mean = cv::mean(cv_cropped_img);
  float global_mean = 0.0, mean_sum_value = 0.0;
  for (int i = 0; i < cv_cropped_img.channels(); i++)
  {
	  mean_sum_value += channel_mean[i];
  }
  global_mean = mean_sum_value / img_channels;
  //LOG(INFO) << "Using opencv to load data ----------------------------------------------";
  Dtype* transformed_data = transformed_blob->mutable_cpu_data();
  int top_index;
  // 遍历像素点
  for (int h = 0; h < height; ++h) 
  {
    const uchar* ptr = cv_cropped_img.ptr<uchar>(h);
    int img_index = 0;
    for (int w = 0; w < width; ++w) 
	{
      for (int c = 0; c < img_channels; ++c) 
	  {
        top_index = (c * height + h) * width + w;
        Dtype pixel = static_cast<Dtype>(ptr[img_index++]);
        transformed_data[top_index] = (pixel * 1.0 - global_mean) * scale;
      }
    }
  }
  /
  // ------------------------------------原有代码--------------------------------------------
  //Dtype* transformed_data = transformed_blob->mutable_cpu_data();
  //int top_index;
  //for (int h = 0; h < height; ++h) {
  //  const uchar* ptr = cv_cropped_img.ptr<uchar>(h);
  //  int img_index = 0;
  //  for (int w = 0; w < width; ++w) {
  //    for (int c = 0; c < img_channels; ++c) {
  //      if (do_mirror) {
  //        top_index = (c * height + h) * width + (width - 1 - w);
  //      } else {
  //        top_index = (c * height + h) * width + w;
  //      }
  //      // int top_index = (c * height + h) * width + w;
  //      Dtype pixel = static_cast<Dtype>(ptr[img_index++]);
  //      if (has_mean_file) {
  //        int mean_index = (c * img_height + h_off + h) * img_width + w_off + w;
  //        transformed_data[top_index] =
  //          (pixel - mean[mean_index]) * scale;
  //      } else {
  //        if (has_mean_values) {
  //          transformed_data[top_index] =
  //            (pixel - mean_values_[c]) * scale;
  //        } else {
  //          transformed_data[top_index] = pixel * scale;
  //        }
  //      }
  //    }
  //  }
  //}

Step2：总结

1. 基本就是找到数据预处理操作的地方，修改 / 重写相应代码
2. 同样的，也可以将预处理做完再进行数据转换，不需要修改代码，看个人喜好
3. resize操作尽可能提前做完，不然拖慢效果拔群
4. 尽可能减少I/O操作，拖慢效果拔群

Step3：修改源码下载地址

下载地址