caffe如何自定义网络以及自定义层（python）（五）

来源：互联网发布：mac桌面图标怎么隐藏编辑：程序博客网时间：2024/04/30 03:40

前面铺垫了这么多，终于到主题了。

先写一个配置文件conv.protxt

layer {  name: 'MyPythonLayer'  type: 'Python'  top: 'output'  python_param {    module: 'mypythonlayer'    layer: 'MyLayer'    param_str: "{\'data_dir\':\'../../images\',\'num\': 100}"  }}

然后就是这个python模块怎么写？mypythonlayer.py

#!/usr/bin/env python#coding=utf-8import sys import osimport numpy as npimport os.path as ospimport matplotlib.pyplot as pltimport pylabfrom copy import copycaffe_root = '/home/x/git/caffe/'sys.path.insert(0, caffe_root + 'python')import caffeimport numpy as npimport yamlimport numpy as npfrom PIL import Imageclass MyLayer(caffe.Layer):    def setup(self, bottom, top):        params = eval(self.param_str)self.data_dir = params['data_dir']        self.num = yaml.load(self.param_str)["num"]self.cat  = '{}/{}.jpg'.format(self.data_dir,'cat')print "Image:",self.cat        print "Parameter num : ", self.num    def reshape(self, bottom, top):        im = Image.open(self.cat)        in_ = np.array(im, dtype=np.float32)        in_ = in_[:,:,::-1]        self.data = in_.transpose((2,0,1))top[0].reshape(1, *self.data.shape)    def forward(self, bottom, top):top[0].data[...] = self.dataprint "forward=============================>cat"print self.dataprint type(top[0])print dir(top[0])print type(top[0].data)#print help(top[0].data)print type(top[0].data[...])    def backward(self, top, propagate_down, bottom):        pass

里面包括几个基本的设置，然后就是参数的读取，我这里添加了两个参数，路径和一个num参数。

基本实现里面有参数读取和图片读取的简单实现。

接下来调用test_python.py

#!/usr/bin/env python#coding=utf-8import numpy as npimport matplotlib.pyplot as pltfrom PIL import Imagecaffe_root = '/home/x/git/caffe/'import syssys.path.insert(0, caffe_root + 'python')import caffeimport mypythonlayersys.path.append("/home/x/git/caffe/examples/test_layer_lb/mypython")caffe.set_device(0)caffe.set_mode_gpu()print "================================="net = caffe.Net('/home/x/git/caffe/examples/test_layer_lb/mypython/conv.prototxt',caffe.TEST)#print net.blobs.items()#print dir(net.blobs)net.forward()

日志：

WARNING: Logging before InitGoogleLogging() is written to STDERRI1129 19:11:46.731223 16691 net.cpp:49] Initializing net from parameters: state {  phase: TEST}layer {  name: "MyPythonLayer"  type: "Python"  top: "output"  python_param {    module: "mypythonlayer"    layer: "MyLayer"    param_str: "{\'data_dir\':\'../../images\',\'num\': 100}"  }}I1129 19:11:46.731266 16691 layer_factory.hpp:77] Creating layer MyPythonLayerI1129 19:11:46.731314 16691 net.cpp:91] Creating Layer MyPythonLayerI1129 19:11:46.731322 16691 net.cpp:399] MyPythonLayer -> outputI1129 19:11:46.744997 16691 net.cpp:141] Setting up MyPythonLayerI1129 19:11:46.745023 16691 net.cpp:148] Top shape: 1 3 360 480 (518400)I1129 19:11:46.745028 16691 net.cpp:156] Memory required for data: 2073600I1129 19:11:46.745035 16691 net.cpp:219] MyPythonLayer does not need backward computation.I1129 19:11:46.745039 16691 net.cpp:261] This network produces output outputI1129 19:11:46.745045 16691 net.cpp:274] Network initialization done.=================================Image: ../../images/cat.jpgParameter num :  100forward=============================>cat[[[  49.   50.   47. ...,   30.   36.   45.]  [  51.   52.   48. ...,   32.   37.   45.]  [  51.   51.   49. ...,   27.   35.   42.]  ...,   [  15.   13.   13. ...,  156.  162.  164.]  [  19.   20.   22. ...,  167.  158.  167.]  [  25.   25.   27. ...,  188.  157.  158.]] [[  57.   58.   55. ...,   77.   87.   94.]  [  57.   58.   56. ...,   82.   89.   96.]  [  57.   57.   57. ...,   88.   95.   99.]  ...,   [  47.   46.   47. ...,  173.  179.  181.]  [  53.   56.   58. ...,  184.  175.  184.]  [  62.   64.   66. ...,  205.  174.  175.]] [[  26.   27.   25. ...,   45.   50.   56.]  [  26.   27.   25. ...,   50.   52.   58.]  [  26.   26.   26. ...,   52.   55.   60.]  ...,   [  36.   32.   36. ...,  182.  188.  190.]  [  42.   42.   44. ...,  193.  184.  193.]  [  46.   49.   51. ...,  214.  183.  184.]]]<class 'caffe._caffe.Blob'>['__class__', '__delattr__', '__dict__', '__doc__', '__format__', '__getattribute__', '__hash__', '__init__', '__module__', '__new__', '__reduce__', '__reduce_ex__', '__repr__', '__setattr__', '__sizeof__', '__str__', '__subclasshook__', '__weakref__', 'channels', 'count', 'data', 'diff', 'height', 'num', 'reshape', 'shape', 'width']<type 'numpy.ndarray'><type 'numpy.ndarray'>

本人水平有点菜，开始一看python的protxt里面的参数有点懵逼，如果看不懂里面的某个东西直接写个文件调试，慢慢就懂了。

a="{\'sbdd_dir\': \'../data/sbdd/dataset\', \'seed\': 1337, \'split\': \'train\', \'mean\': (104.00699, 116.66877, 122.67892)}"params = eval(a)print type(())print params.items()params_dir=params['sbdd_dir']print params_dir

0 0