tf.stop_gradient(), tf.clip_by_value(), tf.trainable_variables(), tf.variables_intializer() の実装に関するメモ

Sample code: tf.stop_gradient()

tf.reset_default_graph()

with tf.variable_scope('test'):
  x = tf.get_variable('x', shape = (), initializer = tf.constant_initializer())

a = 0.5 * x
a = tf.stop_gradient(a)   # check
f = (x - a - 2.0) ** 2
opt = tf.train.GradientDescentOptimizer(learning_rate = 0.1).minimize(f)

with tf.Session() as sess:
  sess.run(tf.global_variables_initializer())

  print('initial')
  print (sess.run(x))
  print (sess.run(f))

  sess.run(opt)
  print('-' * 10)
  print('1st operation')
  print (sess.run(x))
  print (sess.run(f))

  sess.run(opt)
  print('-' * 10)
  print('2nd operation')
  print (sess.run(x))
  print (sess.run(f))

Output results

1. with stop gradient

2. without stop gradient

Sample code: tf.clip_by_value()

​tf.reset_default_graph()

with tf.variable_scope('test_1'):
  init_const = tf.constant_initializer(value = 0.0, dtype = tf.float32)
  x = tf.get_variable('x', shape = [], initializer = init_const)
  y = tf.get_variable('y', shape = [], initializer = init_const)

  y_clipped = tf.clip_by_value(y, 0.0, 2.0)
  f = (x - 2.0)**2 + (y_clipped - 3)**2

  t_vars = tf.trainable_variables('test_1')
  opt = tf.train.GradientDescentOptimizer(0.3).minimize(f, var_list = t_vars)

with tf.Session() as sess:
  sess.run(tf.global_variables_initializer())

  print ('initial')
  print ('x: {:.4f}'.format(sess.run(x)))
  print ('y: {:.4f}'.format(sess.run(y)))
  print ('y_clipped: {:.4f}'.format(sess.run(y_clipped)))
  print ('f: {:.4f}'.format(sess.run(f)))

  for i in range(5):
    sess.run(opt)
    print ('-' * 15)
    print ('iteration: {}'.format(i + 1))
    print ('x: {:.4f}'.format(sess.run(x)))
    print ('y: {:.4f}'.format(sess.run(y)))
    print ('y_clipped: {:.4f}'.format(sess.run(y_clipped)))
    print ('f: {:.4f}'.format(sess.run(f)))

Output results

Sample code: tf.trainable_variables()

tf.reset_default_graph()

with tf.variable_scope('var_1'):
  init_const = tf.constant_initializer(value = 0.0, dtype = tf.float32)
  x = tf.get_variable('x', shape = [], initializer = init_const)

with tf.variable_scope('var_2'):
  init_const = tf.constant_initializer(value = 0.0, dtype = tf.float32)
  y = tf.get_variable('y', shape = [], initializer = init_const)

  f = (x - 2.0)**2 + (y - 3)**2

  t_vars_1 = tf.trainable_variables('var_1')
  t_vars_2 = tf.trainable_variables('var_2')
  t_vars_all = t_vars_1 + t_vars_2
  opt = tf.train.GradientDescentOptimizer(0.3).minimize(f, var_list = t_vars_1)

with tf.Session() as sess:
  sess.run(tf.global_variables_initializer())

  print ('initial')
  print ('x: {:.4f}'.format(sess.run(x)))
  print ('y: {:.4f}'.format(sess.run(y)))
  print ('f: {:.4f}'.format(sess.run(f)))

  for i in range(5):
    sess.run(opt)
    print ('-' * 15)
    print ('iteration: {}'.format(i + 1))
    print ('x: {:.4f}'.format(sess.run(x)))
    print ('y: {:.4f}'.format(sess.run(y)))
    print ('f: {:.4f}'.format(sess.run(f)))

Output results

Sample code: tf.variables_initializer()

tf.reset_default_graph()

with tf.variable_scope('test'):
  init_const_1 = tf.constant_initializer(value = 0.0, dtype = tf.float32)
  init_const_2 = tf.constant_initializer(value = 5.0, dtype = tf.float32)
  x = tf.get_variable('x', shape = [], initializer = init_const_1)
  y = tf.get_variable('y', shape = [2, 2], initializer = init_const_2)

with tf.Session() as sess:

  u_vars = tf.report_uninitialized_variables()
  print (sess.run(u_vars))

  test_vars = tf.trainable_variables('test')

  for var in test_vars:
    sess.run(tf.variables_initializer([var]))

  print (sess.run(x))
  print (sess.run(y))

Output results