import tensorflow as tf
import argparse
import os
parser = argparse.ArgumentParser(description='')
parser.add_argument("--checkpoint_path", default='./init/model.ckpt', help="restore ckpt") #原参数路径
parser.add_argument("--new_checkpoint_path", default='./ckpt/', help="path_for_new ckpt") #新参数保存路径
parser.add_argument("--add_prefix", default='deeplab_v2/', help="prefix for addition") #新参数名称中加入的前缀名
args = parser.parse_args()
def main():
if not os.path.exists(args.new_checkpoint_path):
os.makedirs(args.new_checkpoint_path)
with tf.Session() as sess:
new_var_list=[] #新建一个空列表存储更新后的Variable变量
for var_name, shape in tf.contrib.framework.list_variables(args.checkpoint_path): #得到checkpoint文件中所有的参数（名字，形状）元组
var = tf.contrib.framework.load_variable(args.checkpoint_path, var_name) #得到上述参数的值
new_name = var_name
new_name = args.add_prefix + new_name #在这里加入了名称前缀，大家可以自由地作修改
print("new_name=",new_name, "shape=",shape)
#除了修改参数名称，还可以修改参数值（var）
print('Renaming %s to %s.' % (var_name, new_name))
renamed_var = tf.Variable(var, name=new_name) #使用加入前缀的新名称重新构造了参数
new_var_list.append(renamed_var) #把赋予新名称的参数加入空列表
print('starting to write new checkpoint !')
saver = tf.train.Saver(var_list=new_var_list) #构造一个保存器
sess.run(tf.global_variables_initializer()) #初始化一下参数（这一步必做）
model_name = 'deeplab_resnet_altered' #构造一个保存的模型名称
checkpoint_path = os.path.join(args.new_checkpoint_path, model_name) #构造一下保存路径
saver.save(sess, checkpoint_path) #直接进行保存
print("done !")
if __name__ == '__main__':
main()
=====================================================
下面这个，就是查看ckpt文件里面所有变量的名称：
from tensorflow.python import pywrap_tensorflow
import os
checkpoint_path = "./init/model.ckpt"
reader = pywrap_tensorflow.NewCheckpointReader(checkpoint_path)
var_to_shape_map = reader.get_variable_to_shape_map()
f = open('./ckpt/ckpt_values.txt','w')
for key in var_to_shape_map:
print("tensor_name: ", key)
print(reader.get_tensor(key)) # Remove this is you want to print only variable names
f.write(key)
f.write("\n")
#f.write(str(reader.get_tensor(key)))
#f.write("\n")
f.close()
