with tf.Graph().as_default() as g: 
  #建立model
  images, labels = cifar10.inputs(eval_data=eval_data) 
  logits = cifar10.inference(images) 
  top_k_op = tf.nn.in_top_k(logits, labels, 1) 
  #從ckpt中恢復變量
  sess = tf.Session()
  saver = tf.train.Saver() #saver = tf.train.Saver(...variables...) # 恢復部分變量時，只需要在Saver里指定要恢復的變量
  save_path = 'ckpt的路徑'
  saver.restore(sess, save_path) # 從ckpt中恢復變量

#使用NewCheckpointReader來讀取ckpt里的變量
from tensorflow.python import pywrap_tensorflow
checkpoint_path = os.path.join(model_dir, "model.ckpt")
reader = pywrap_tensorflow.NewCheckpointReader(checkpoint_path) #tf.train.NewCheckpointReader
var_to_shape_map = reader.get_variable_to_shape_map()
for key in var_to_shape_map:
  print("tensor_name: ", key)
  #print(reader.get_tensor(key))

#使用print_tensors_in_checkpoint_file打印ckpt里的內(nèi)容
from tensorflow.python.tools.inspect_checkpoint import print_tensors_in_checkpoint_file

print_tensors_in_checkpoint_file(file_name, #ckpt文件名字
                 tensor_name, # 如果為None,則默認為ckpt里的所有變量
                 all_tensors, # bool 是否打印所有的tensor，這里打印出的是tensor的值，一般不推薦這里設置為False
                 all_tensor_names) # bool 是否打印所有的tensor的name
#上面的打印ckpt的內(nèi)部使用的是pywrap_tensorflow.NewCheckpointReader所以，掌握NewCheckpointReader才是王道

from tensorflow.python.tools import freeze_graph

freeze_graph(input_graph, #=some_graph_def.pb
       input_saver, 
       input_binary, 
       input_checkpoint, #=model.ckpt
       output_node_names, #=softmax
       restore_op_name, 
       filename_tensor_name, 
       output_graph, #='./tmp/frozen_graph.pb'
       clear_devices, 
       initializer_nodes, 
       variable_names_whitelist='', 
       variable_names_blacklist='', 
       input_meta_graph=None, 
       input_saved_model_dir=None, 
       saved_model_tags='serve', 
       checkpoint_version=2)
#freeze_graph_test.py講述了怎么使用freeze_grapg。