Tensorflow---VariableScope and NameScope

1. VariableScope and NameScope

1.1 variable scope

tf.variable_scope()是用来实现变量共享 : 可以作用于var和op.
e.g.

with tf.variable_scope("v_scope_1"):
    v1 = tf.get_variable("variable1", [6], initializer=tf.constant_initializer(1.0))
    v2 = tf.get_variable("variable2", [6], initializer=tf.constant_initializer(2.0))
print v1.name
print v2.name

v_scope_1/variable1:0
v_scope_1/variable2:0

接口详细分析
tf.variable_scope()

__init__(
    name_or_scope,
    default_name=None,
    values=None,
    initializer=None,
    regularizer=None,
    caching_device=None,
    partitioner=None,
    custom_getter=None,
    reuse=None,
    dtype=None,
    use_resource=None,
    constraint=None,
    auxiliary_name_scope=True
)

name_or_scope:
可以是name也可以是已经创建的scope，用法如下:

initializer:
在scope中使用本参数初始化var

with tf.variable_scope("v_scope_1", initializer=tf.constant_initializer(6.)):
    v1 = tf.get_variable("variable1", [6])

v1=[6. 6. 6. 6. 6. 6.]

reuse:
在var创建后，share这个var

def foo():
  with tf.variable_scope("foo", reuse=tf.AUTO_REUSE):
    v = tf.get_variable("v", [1])
  return v

v1 = foo()  # Creates v.
v2 = foo()  # Gets the same, existing v.
assert v1 == v2

with tf.variable_scope("foo"):
    v = tf.get_variable("v", [1])
with tf.variable_scope("foo", reuse=True):
    v1 = tf.get_variable("v", [1])
assert v1 == v

with tf.variable_scope("foo") as scope:
    v = tf.get_variable("v", [1])
    scope.reuse_variables()
    v1 = tf.get_variable("v", [1])
assert v1 == v

auxiliary_name_scope
暂时未知

1.2 name scope

tf.name_scope()是op的名称空间 : 可以作用与op.

2. Save and Restore

2.1 简单的model save 和 restore

model save

with tf.variable_scope("v_scope_1", initializer=tf.constant_initializer(6.)):
    v1 = tf.get_variable("variable1", [1])
    
with tf.name_scope("v_assign_op_1"):
    v1 = tf.assign(v1, [8.0])
    
saver = tf.train.Saver()    
    
with tf.Session() as sess:
    vars_init_op = tf.global_variables_initializer()
    sess.run(vars_init_op)
    r1 = sess.run(v1)
    print r1

model restore : v1在上面的代码中初始化为6.，后被assign置为8.并存在model中，然后在下面的代码中读取

with tf.variable_scope("v_scope_1", initializer=tf.constant_initializer(6.)):
    v1 = tf.get_variable("variable1", [1])
    
#with tf.name_scope("v_assign_op_1"):
#    v1 = tf.assign(v1, [8.0])
    
saver = tf.train.Saver()    
    
with tf.Session() as sess:
    #vars_init_op = tf.global_variables_initializer()
    #sess.run(vars_init_op)
    saver.restore(sess, "./model.ckpt")
    r1 = sess.run(v1)
    print r1

2.2 在save和restore时改名，部分restore

挑选var保存 : 只保存var_list中的变量

var_list = []

with tf.variable_scope("v_scope_1", initializer=tf.constant_initializer(6.)):
    v1 = tf.get_variable("variable1", [1])
    v2 = tf.get_variable("variable2", [1])

var_list.append(v1)
    
with tf.name_scope("v_assign_op_1"):
    v1 = tf.assign(v1, [8.0])
    v2 = tf.assign(v2, [8.0])
    
saver = tf.train.Saver(var_list)    
    
with tf.Session() as sess:
    vars_init_op = tf.global_variables_initializer()
    sess.run(vars_init_op)
    r1, r2 = sess.run([v1, v2])
    print r1, r2
    saver.save(sess, "./model.ckpt")

[8.] [8.]

---

挑选var读取 : 注意读取时name和scope都要对应上

var_list = []

with tf.variable_scope("v_scope_1", initializer=tf.constant_initializer(6.)):
    v1 = tf.get_variable("variable1", [1])
    v2 = tf.get_variable("variable2", [1])

var_list.append(v1)
    
# with tf.name_scope("v_assign_op_1"):
#     v1 = tf.assign(v1, [8.0])
#     v2 = tf.assign(v2, [8.0])
    
saver = tf.train.Saver(var_list)    
    
with tf.Session() as sess:
    vars_init_op = tf.global_variables_initializer()
    sess.run(vars_init_op)
    saver.restore(sess, "./model.ckpt")
    r1, r2 = sess.run([v1, v2])
    print r1, r2

[8.] [6.]

---

通过dict保存、读取

3. npy等其它save方式