Tensorflow: how to save/restore a model?

fantomas

After you train a model in Tensorflow:

• How do you save the trained model?
• How do you later restore this saved model?
  

Save the model:

1. <font color="#101094" face="Consolas, Menlo, Monaco, Lucida Console, Liberation Mono, DejaVu Sans Mono, Bitstream Vera Sans Mono, Courier New, monospace, sans-serif"><span style="font-size: 13px; white-space: pre;">import tensorflow as tf
   
2. 
   
3. #Prepare to feed input, i.e. feed_dict and placeholders
   
4. w1 = tf.placeholder("float", name="w1")
   
5. w2 = tf.placeholder("float", name="w2")
   
6. b1= tf.Variable(2.0,name="bias")
   
7. feed_dict ={w1:4,w2:8}
   
8. 
   
9. #Define a test operation that we will restore
   
10. w3 = tf.add(w1,w2)
    
11. w4 = tf.multiply(w3,b1,name="op_to_restore")
    
12. sess = tf.Session()
    
13. sess.run(tf.global_variables_initializer())
    
14. 
    
15. #Create a saver object which will save all the variables
    
16. saver = tf.train.Saver()
    
17. 
    
18. #Run the operation by feeding input
    
19. print sess.run(w4,feed_dict)
    
20. #Prints 24 which is sum of (w1+w2)*b1
    
21. 
    
22. #Now, save the graph
    
23. saver.save(sess, 'my_test_model',global_step=1000)</span></font>

复制代码

Restore the model:

1. <div class="blockcode"><blockquote><blockquote>import tensorflow as tf

复制代码

1. import tensorflow as tf
   
2. 
   
3. sess=tf.Session()   
   
4. #First let's load meta graph and restore weights
   
5. saver = tf.train.import_meta_graph('my_test_model-1000.meta')
   
6. saver.restore(sess,tf.train.latest_checkpoint('./'))
   
7. 
   
8. 
   
9. # Access saved Variables directly
   
10. print(sess.run('bias:0'))
    
11. # This will print 2, which is the value of bias that we saved
    
12. 
    
13. 
    
14. # Now, let's access and create placeholders variables and
    
15. # create feed-dict to feed new data
    
16. 
    
17. graph = tf.get_default_graph()
    
18. w1 = graph.get_tensor_by_name("w1:0")
    
19. w2 = graph.get_tensor_by_name("w2:0")
    
20. feed_dict ={w1:13.0,w2:17.0}
    
21. 
    
22. #Now, access the op that you want to run.
    
23. op_to_restore = graph.get_tensor_by_name("op_to_restore:0")
    
24. 
    
25. print sess.run(op_to_restore,feed_dict)
    
26. #This will print 60 which is calculated

复制代码

This and some more advanced use-cases have been explained very well here.

AIcomes手机认证

有用，帮助很大