caffe python layer

Layer實際上定義了Layer的基本操做，即初始化層、前向傳播和反向傳播。在前向傳播中根據bottom blob獲得top blob，反向傳播則根據top反傳到bottom。並且在前傳的時候還能夠計算loss，通常來講只有最後一層纔會計算loss，雖然每一個層都有計算loss的功能。Layer類在沒有實現GPU前傳和反傳的時候會自動使用CPU的實現。下面給出Layer類的具體介紹。

layer {
  name: "data"
  type: "Python"
  top: "data"
  top: "label"
  python_param {
    module: "voc_layers"   #模塊名，必須在PYTHONPATH中
    layer: "SBDDSegDataLayer" #模塊中的類名
    param_str: "{\'sbdd_dir\': \'../data/sbdd/dataset\', \'seed\': 1337, \'split\': \'train\', \'mean\': (104.00699, 116.66877, 122.67892)}" #這裏面的東西都會在setup以前被放在PythonLayer對象的屬性中，在setup中解析這些參數，在forward和backward等中使用。
  }
}

 

class VOCSegDataLayer(caffe.Layer):

 

     def  setup( self , bottom, top):

   

         # config

         params  =  eval ( self .param_str)

         self .voc_dir  =  params[ 'voc_dir' ]         #是否是以爲挺突兀的？爲何在include裏的PythonLayer類中沒有定義voc_dir這些屬性，VOCSegDataLayer類中也沒有定義這些屬性，此處卻表示VOCSegDataLayer有這些屬性呢？由於是自動放進去的，在python_layer.hpp中有相關代碼

         self .split  =  params[ 'split' ]

         self .mean  =  np.array(params[ 'mean' ])

         self .random  =  params.get( 'randomize' ,  True )

         self .seed  =  params.get( 'seed' ,  None )

 

         # two tops: data and label

         if  len (top) ! =  2 :

             raise  Exception( "Need to define two tops: data and label." )

         # data layers have no bottoms

         if  len (bottom) ! =  0 :

             raise  Exception( "Do not define a bottom." )

 

         # load indices for images and labels

         split_f   =  '{}/ImageSets/Segmentation/{}.txt' . format ( self .voc_dir,

                 self .split)

         self .indices  =  open (split_f,  'r' ).read().splitlines()

         self .idx  =  0

 

         # make eval deterministic

         if  'train'  not  in  self .split:

             self .random  =  False

 

         # randomization: seed and pick

         if  self .random:

             random.seed( self .seed)

             self .idx  =  random.randint( 0 ,  len ( self .indices) - 1 )

 

 

     def  reshape( self , bottom, top):

         # load image + label image pair

         self .data  =  self .load_image( self .indices[ self .idx])

         self .label  =  self .load_label( self .indices[ self .idx])

         # reshape tops to fit (leading 1 is for batch dimension)

         top[ 0 ].reshape( 1 ,  * self .data.shape)

         top[ 1 ].reshape( 1 ,  * self .label.shape)

 

 

     def  forward( self , bottom, top):

         # assign output

         top[ 0 ].data[...]  =  self .data

         top[ 1 ].data[...]  =  self .label

 

         # pick next input

         if  self .random:

             self .idx  =  random.randint( 0 ,  len ( self .indices) - 1 )

         else :

             self .idx  + =  1

             if  self .idx  = =  len ( self .indices):

                 self .idx  =  0

 

 

     def  backward( self , top, propagate_down, bottom):

         pass

 

 

     def  load_image( self , idx):

         """

         Load input image and preprocess for Caffe:

         - cast to float

         - switch channels RGB -> BGR

         - subtract mean

         - transpose to channel x height x width order

         """

         im  =  Image. open ( '{}/JPEGImages/{}.jpg' . format ( self .voc_dir, idx))

         in_  =  np.array(im, dtype = np.float32)

         in_  =  in_[:,:,:: - 1 ]

         in_  - =  self .mean

         in_  =  in_.transpose(( 2 , 0 , 1 ))

         return  in_

 

 

https://chrischoy.github.io/research/caffe-python-layer/

http://blog.csdn.net/thesby/article/details/51264439

https://www.cnblogs.com/fanhaha/p/7247839.html

這兩個博客講的好