利用python输出caffemodel的权值（weights）参数至txt文档中。

#-*- coding: UTF-8 -*-
#!/usr/bin/env python  
  
# 引入“咖啡”  
import caffe  
  
import numpy as np  
root = '/home/x306/xn/caffe-master/'  
# 使输出的参数完全显示  
# 若没有这一句，因为参数太多，中间会以省略号“……”的形式代替  
np.set_printoptions(threshold='nan')  
  
# deploy文件  
MODEL_FILE = root + 'examples/facetestquestions/deploy.prototxt'  
# 预先训练好的caffe模型  
PRETRAIN_FILE = root + 'examples/facetestquestions/model/mnin_plant_iter_30000.caffemodel'  
  
# 保存参数的文件  
params_txt = root + 'examples/facetestquestions/params.txt'  
pf = open(params_txt, 'w')  
  
# 让caffe以测试模式读取网络参数  
net = caffe.Net(MODEL_FILE, PRETRAIN_FILE, caffe.TEST)  
  
# 遍历每一层  
for param_name in net.params.keys():  
    # 权重参数  
    weight = net.params[param_name][0].data  
    # 偏置参数  
    bias = net.params[param_name][1].data  
  
    # 该层在prototxt文件中对应“top”的名称  
    pf.write(param_name)  
    pf.write('\n')  
  
    # 写权重参数  
    pf.write('\n' + param_name + '_weight:\n\n')  
    # 权重参数是多维数组，为了方便输出，转为单列数组  
    weight.shape = (-1, 1)  
  
    for w in weight:  
        pf.write('%ff, ' % w)  
  
    # 写偏置参数  
    pf.write('\n\n' + param_name + '_bias:\n\n')  
    # 偏置参数是多维数组，为了方便输出，转为单列数组  
    bias.shape = (-1, 1)  
    for b in bias:  
        pf.write('%ff, ' % b)  
  
    pf.write('\n\n')  
  
pf.close