如何在caffe中增加实时数据增强？

东方耀

如何在caffe中增加实时数据增强？



数据层type: "ImageData"

网络结构的prototxt

transform_param {
      mean_value: 127.5
      mean_value: 127.5
      mean_value: 127.5
      scale: 0.0078125
      mirror: true     # 代表随机的左右翻转
      contrast_brightness_adjustment: true    # 对比度调节
        min_contrast: 0.8    # 最小的对比度乘子(min alpha)，默认0.8
        max_contrast: 1.2    # 最大对比度乘子(max alpha)，默认1.2
        max_color_shift: 20
        smooth_filtering: true
        max_smooth: 6       # 最大平滑乘子，默认6   进行高斯平滑
        apply_probability: 0.5
  }
  

报错："caffe.TransformationParameter" has no field named "contrast_brightness_adjustment".


在MY_caffe_root/src/caffe/proto/caffe.proto  修改：
// Message that stores parameters used to apply transformation
// to the data layer's data
message TransformationParameter {
  // For data pre-processing, we can do simple scaling and subtracting the
  // data mean, if provided. Note that the mean subtraction is always carried
  // out before scaling.
  optional float scale = 1 [default = 1];
  // Specify if we want to randomly mirror data.
  optional bool mirror = 2 [default = false];
  // Specify if we would like to randomly crop an image.
  optional uint32 crop_size = 3 [default = 0];
  // mean_file and mean_value cannot be specified at the same time
  optional string mean_file = 4;
  // if specified can be repeated once (would subtract it from all the channels)
  // or can be repeated the same number of times as channels
  // (would subtract them from the corresponding channel)
  repeated float mean_value = 5;
  // Force the decoded image to have 3 color channels.
  optional bool force_color = 6 [default = false];
  // Force the decoded image to have 1 color channels.
  optional bool force_gray = 7 [default = false];


  // Image augmentation  dfy加的 数据增强的
  // Specify the range of scaling factor for doing resizing
  optional float min_scaling_factor = 8 [default = 0.75];
  optional float max_scaling_factor = 9 [default = 1.50];
  // Specify the angle for doing rotation
  optional uint32 max_rotation_angle = 10 [default = 0];
  // Specify the contrast, brightness, smooth and color shift for augmentation
  optional bool contrast_brightness_adjustment = 11 [default = false];
  optional bool smooth_filtering = 12 [default = false];
  optional float min_contrast = 14 [default = 0.8];
  optional float max_contrast = 15 [default = 1.2];
  optional uint32 max_brightness_shift = 16 [default = 5];
  optional float max_smooth = 17 [default = 6];
  optional uint32 max_color_shift = 20 [default = 0];

  // Min side resizing, keep aspect ratio
  optional uint32 min_side_min = 13 [default = 0];
  optional uint32 min_side_max = 21 [default = 0];
  optional uint32 min_side = 22 [default = 0];

  optional float apply_probability = 18 [default = 0.5];
  optional bool debug_params = 19 [default = false];
}


修改源码：
MY_caffe_root/src/caffe/data_transformer.cpp


阅读源码“data_transformer.cpp”可知，执行顺序是mean(file)、mean(value)、crop、mirror、scale。

如有其它特殊的预处理可在此添加代码





之后  make clean

make all -j8
make test -j8
make pycaffe -j8

第七步：拷贝编译后的libcaffe.so至caffe的虚拟环境

在caffe源码项目下/python目录下 打开终端：
cp -rf caffe/ ~/miniconda3/envs/py2_caffe_source/lib/python2.7/
cp -rf ../.build_release/lib/* ~/miniconda3/envs/py2_caffe_source/lib/

东方耀

数据增广代码参考caffe-augmentation
https://github.com/twtygqyy/caffe-augmentation
Realtime data augmentation is implemented within the ImageData layer.

东方耀

待研究：减少显存消耗的，代码来自caffe-Mem-Op，主要代码在net.cpp里。
https://github.com/yjxiong/caffe

东方耀

使用caffe-augmentation时怎么修改original caffe吗?需要修改哪些？