[TF] How to save and load Tensorflow learning parameters

Use ** tf.train.Saver ** to save and load parameters learned in Tensorflow.

Save

When saving, use the ** save ** method of the created saver class.

python

saver = tf.train.Saver()

Some processing
 
#Save
saver.save(sess, "model.ckpt")

Saving can be at the end of learning or at the middle of learning.

Read

When reading, use the ** restore ** method of the created saver class. Since session is required, load it after creating session. When running on ipython, create a session with tf.InteractiveSession (), usually tf.Session ().

python

sess=tf.InteractiveSession()

saver.restore(sess, "model.ckpt")

The state of actually saving and loading is shown below.

The flow is as follows.

1. 1. Modeling
2. Learning
3. 3. Save parameters to another variable for later comparison
4. Save parameters to file ５．Session Close
5. Session creation
6. Initialization (This is not necessary in the first place. It was intentionally initialized for comparison.)
7. Compare with the saved parameters (This is different because it was initialized one time ago.)
8. Read parameters from file
9. Compare with saved parameters (this matches)
10. Learning

code

python

# # import

# In[1]:

import tensorflow as tf
import numpy as np
from tensorflow.examples.tutorials.mnist import input_data


# # load dataset

# In[2]:

mnist = input_data.read_data_sets("./data/mnist/", one_hot=True) 


# # build model

# In[3]:

def mlp(x, output_dim, reuse=False):
        
    w1 = tf.get_variable("w1", [x.get_shape()[1], 1024], initializer=tf.random_normal_initializer())
    b1 = tf.get_variable("b1", [1024], initializer=tf.constant_initializer(0.0))
    w2 = tf.get_variable("w2", [1024, output_dim], initializer=tf.random_normal_initializer())
    b2 = tf.get_variable("b2", [output_dim], initializer=tf.constant_initializer(0.0))
    
    fc1 = tf.nn.relu(tf.matmul(x, w1) + b1)
    fc2 = tf.matmul(fc1, w2) + b2

    return fc2, [w1, b1, w2, b2]

def slp(x, output_dim):
    w1 = tf.get_variable("w1", [x.get_shape()[1], output_dim], initializer=tf.random_normal_initializer())
    b1 = tf.get_variable("b1", [output_dim], initializer=tf.constant_initializer(0.0))
    
    fc1 = tf.nn.relu(tf.matmul(x, w1) + b1)
    return fc1, [w1, b1]

n_batch = 32
n_label = 10
n_train = 10000
imagesize = 28
learning_rate = 0.5

x_node = tf.placeholder(tf.float32, shape=(n_batch, imagesize*imagesize))
y_node = tf.placeholder(tf.float32, shape=(n_batch, n_label))

with tf.variable_scope("MLP") as scope:
    out_m, theta_m = mlp(x_node, n_label)
           
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(out_m, y_node))
opt  = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss)
tr_pred = tf.nn.softmax(out_m)

test_data = mnist.test.images
test_labels = mnist.test.labels
tx = tf.constant(test_data)
ty_ = tf.constant(test_labels)

with tf.variable_scope("MLP") as scope:
    scope.reuse_variables()
    ty, _ = mlp(tx, n_label)
    
te_pred = tf.nn.softmax(ty) 


# In[4]:

def accuracy(y, y_):
    return 100.0 * np.sum(np.argmax(y, 1) == np.argmax(y_, 1)) / y.shape[0]


# In[5]:

saver = tf.train.Saver()

sess=tf.InteractiveSession()

init = tf.initialize_all_variables()
sess.run(init)


# In[6]:

for step in xrange(n_train):
    bx, by = mnist.train.next_batch(n_batch)
    feed_dict = {x_node: bx, y_node: by}
    _, _loss, _tr_pred = sess.run([opt, loss, tr_pred], feed_dict=feed_dict)
    if step % 500 == 0:
        _accuracy = accuracy(_tr_pred, by)
        print 'step = %d, loss=%.2f, accuracy=%.2f' % (step, _loss, _accuracy)

print 'test accuracy=%.2f' % accuracy(te_pred.eval(), test_labels)


# In[8]:

old_theta_m = [ p.eval() for p in theta_m] # for comparing


# In[9]:

saver.save(sess, "model.ckpt")


# In[10]:

sess.close()


# In[11]:

sess=tf.InteractiveSession()

# for clear
init = tf.initialize_all_variables()
sess.run(init)


# In[12]:

for prm, prm_o in zip(theta_m, old_theta_m):
    p1 = prm.eval()
    p2 = prm_o
    print np.sum(p1 != p2) 


# In[13]:

saver.restore(sess, "model.ckpt")


# In[14]:

for prm, prm_o in zip(theta_m, old_theta_m):
    p1 = prm.eval()
    p2 = prm_o
    print np.sum(p1 != p2) 


# In[15]:

print 'test accuracy=%.2f' % accuracy(te_pred.eval(), test_labels)


# In[16]:

for step in xrange(n_train):
    bx, by = mnist.train.next_batch(n_batch)
    feed_dict = {x_node: bx, y_node: by}
    _, _loss, _tr_pred = sess.run([opt, loss, tr_pred], feed_dict=feed_dict)
    if step % 500 == 0:
        _accuracy = accuracy(_tr_pred, by)
        print 'step = %d, loss=%.2f, accuracy=%.2f' % (step, _loss, _accuracy)

print 'test accuracy=%.2f' % accuracy(te_pred.eval(), test_labels)


# In[17]:

sess.close()


# In[ ]:

tf.reset_default_graph()