用python自己写个caffe的层

有时如果需要自己添加层到Caffe当中，这时候需要自己编写py文件

#!/usr/bin/envs python
#coding=utf-8
import sys
caffe_root='/home/dyh/caffe/caffe/'
sys.path.insert(0, caffe_root+'python')
import numpy as np
import caffe
import cv2
import yaml
#类名一定要和配置文件里面的layer相同
class MyLayer(caffe.Layer): #这是一个caffe类,不要变
    #caffe规定按如下方式写
    def setup(self, bottom, top):
        self.num = yaml.load(self.param_str)['num']
        print('Parameter num:',self.num)
    def reshape(self, bottom, top):
        pass
    def forward(self, bottom, top):
        top[0].reshape(*bottom[0].shape)#输出和输入的shape一样
        print(bottom[0].data.shape)
        print(bottom[0].data)
        #测试一下加一个数值能不能行
        top[0].data[...] = bottom[0].data + self.num#输出的data就等于输入的加上自定义层的参数
        print('输出',top[0].data[...])
    def backward(self,top,bottom,propagate_down):
        pass

net = caffe.Net('/home/dyh/eclipse-workspace/MyCaffeLayer/src/conv.prototxt',caffe.TEST)
im = np.array(cv2.imread('/home/dyh/eclipse-workspace/MyCaffeLayer/src/cat.jpg'))
print(im.shape)
#添加一个维度[h,w,c]->[b,h,w,c],
im_input = im[np.newaxis,:,:]
print(im_input.shape)
#要将[b,h,w,c]->[b,c,h,w]要符合caffe的规定
im_input2 = im_input.transpose([0,3,1,2])
print(im_input2.shape)
#*表示zip成一个seq，网络的data层是1,3,100,100,reshape和im_input2一样
net.blobs['data'].reshape(*im_input2.shape)
#数据很多，用...代替
net.blobs['data'].data[...] = im_input2#把数据单发到data层
net.forward()

conv.prototxt的简单内容如下，具体内容如何需要具体实施，这个只是简单的加法操作

name: "convolution"
input: "data"
input_dim: 1
input_dim: 3
input_dim: 100
input_dim: 100
layer {
	name: "conv"
	type: "Convolution"
	bottom: "data"
	top: "conv"
	convolution_param {
		num_output:3
		kernel_size: 5
		stride: 1
		weight_filler {
			type: "gaussian"
			std: 0.01
		}
		bias_filler{
			type: "constant"
			value: 0
		}
	}
}
#自己定义的层按如下格式
layer{
	name:"MyPythonLayer"#名字自己写，后续接的上就行
	type:'Python'#类型一定是这个
	bottom:'conv'
	top:'output'
	python_param{
		module:'mypythonlayer'#此处和自己定义层的py文件名一样
		layer:'MyLayer'#此处要和自己定义的class一样
		param_str:"'num': 21"#这里数字前面要加空格，注意！！因该是提取的时候需要以空格分界
	}
}

输出如下

(496, 700, 3)
(1, 496, 700, 3)
(1, 3, 496, 700)
(1, 3, 492, 696)
[[[[ 11.0702715   11.13561     11.150491   ...  13.735498
     13.558763    13.209033  ]
   [ 11.158289    11.258202    11.289419   ...  13.523567
     13.414061    12.992344  ]
   [ 11.340612    11.440386    11.423347   ...  13.58939
     13.315084    13.156676  ]
   ...
   [  8.85025      8.900326     8.8069105  ...   9.247419
      9.061868     9.001272  ]
   [  8.848199     8.788747     8.691421   ...   9.226819
      9.138367     8.871523  ]
   [  8.832649     8.70563      8.590294   ...   9.1012335
      9.061207     8.858202  ]]

  [[-33.13818    -32.91376    -32.57329    ... -17.359497
    -17.008904   -16.725342  ]
   [-33.628136   -33.54023    -33.262566   ... -17.534294
    -17.166283   -16.70674   ]
   [-33.889587   -33.84192    -33.683243   ... -17.484074
    -17.18139    -16.803131  ]
   ...
   [-58.23536    -58.174156   -58.293545   ... -10.737489
    -10.5319605  -10.538096  ]
   [-58.214077   -58.24072    -58.419518   ... -10.745275
    -10.566773   -10.593955  ]
   [-58.26086    -58.29385    -58.45631    ... -11.090007
    -10.717339   -10.459574  ]]

  [[ -3.7222815   -3.7532597   -3.728013   ...  -2.322114
     -2.5792527   -2.3096333 ]
   [ -3.9120486   -3.9047034   -3.8735025  ...  -2.3396149
     -2.4620175   -2.3884416 ]
   [ -4.085577    -4.0669966   -3.9437149  ...  -2.4459817
     -2.2547388   -2.3984628 ]
   ...
   [ -4.3689103   -4.2448835   -4.1534586  ...  -0.92515635
     -0.8958546   -0.985209  ]
   [ -4.2951756   -4.1776423   -4.202016   ...  -0.96538967
     -1.0222634   -0.97236586]
   [ -4.2042675   -4.1664195   -4.2614     ...  -0.9604343
     -1.1583881   -0.96595764]]]]
('\xe8\xbe\x93\xe5\x87\xba', array([[[[ 32.07027  ,  32.13561  ,  32.15049  , ...,  34.735497 ,
           34.55876  ,  34.209034 ],
         [ 32.158287 ,  32.2582   ,  32.28942  , ...,  34.523567 ,
           34.414062 ,  33.992344 ],
         [ 32.340614 ,  32.440384 ,  32.423347 , ...,  34.58939  ,
           34.315086 ,  34.156677 ],
         ...,
         [ 29.85025  ,  29.900326 ,  29.806911 , ...,  30.24742  ,
           30.061867 ,  30.001272 ],
         [ 29.848198 ,  29.788746 ,  29.691422 , ...,  30.226818 ,
           30.138367 ,  29.871523 ],
         [ 29.83265  ,  29.705631 ,  29.590294 , ...,  30.101234 ,
           30.061207 ,  29.858202 ]],

        [[-12.13818  , -11.913761 , -11.573292 , ...,   3.640503 ,
            3.9910965,   4.274658 ],
         [-12.628136 , -12.54023  , -12.262566 , ...,   3.4657059,
            3.8337173,   4.2932606],
         [-12.889587 , -12.841919 , -12.683243 , ...,   3.5159264,
            3.8186092,   4.196869 ],
         ...,
         [-37.23536  , -37.174156 , -37.293545 , ...,  10.262511 ,
           10.4680395,  10.461904 ],
         [-37.214077 , -37.24072  , -37.419518 , ...,  10.254725 ,
           10.433227 ,  10.406045 ],
         [-37.26086  , -37.29385  , -37.45631  , ...,   9.909993 ,
           10.282661 ,  10.540426 ]],

        [[ 17.277718 ,  17.24674  ,  17.271988 , ...,  18.677887 ,
           18.420748 ,  18.690367 ],
         [ 17.087952 ,  17.095297 ,  17.126497 , ...,  18.660385 ,
           18.537983 ,  18.611559 ],
         [ 16.914423 ,  16.933002 ,  17.056286 , ...,  18.554018 ,
           18.745262 ,  18.601538 ],
         ...,
         [ 16.63109  ,  16.755116 ,  16.846542 , ...,  20.074844 ,
           20.104145 ,  20.014791 ],
         [ 16.704824 ,  16.822357 ,  16.797985 , ...,  20.03461  ,
           19.977737 ,  20.027634 ],
         [ 16.795732 ,  16.83358  ,  16.7386   , ...,  20.039566 ,
           19.841612 ,  20.034042 ]]]], dtype=float32))

可以看到原来输入进来的图片的像素值都已经增加了21