with tf.name_scope(name):
- name_scope: 为了更好地管理变量的命名空间而提出的。比如在
tensorboard
中,因为引入了name_scope
, 我们的Graph
看起来才井然有序。name_scope
对get_variable
创建变量的 name 没有影响,即get_variable
创建的变量不在name_scope
这个命名空间中
with tf.variable_scope(name_or_scope, reuse=None):
- variable_scope: 大部分情况下,跟
tf.get_variable()
配合使用,实现变量共享
的功能- 可通过
tf.get_variable_scope().reuse == True/False
判断参变量是否共享- 当前变量作用域可以用
tf.get_variable_scope()
进行检索并且reuse
标签可以通过调用tf.get_variable_scope().reuse_variables()
设置为True
tf.Variable()
创建同一个 name 的变量(操作名不同),均不会报错,但系统会自动修改 name(实质还是不让共享参变量)tf.get_varible()
创建同一个 name 的变量(操作名不同),均会报错(为了避免无意识的参变量复用造成的错误)variable_scope
中使用 tf.get_variable()
创建变量,并通过 with tf.variable_scope(name_or_scope, reuse=True)
来共享参变量: tf.get_variable
函数将报错。tf.get_variable
操作将创建新的变量,如果同名的变量已经存在,则tf.get_variable
函数将报错。# 下面是定义一个卷积层的通用方式
def conv_relu(input, kernel_shape, bias_shape):
# Create variable named "weights".
weights = tf.get_variable("weights", kernel_shape,
initializer=tf.random_normal_initializer())
# Create variable named "biases".
biases = tf.get_variable("biases", bias_shape,
initializer=tf.constant_intializer(0.0))
conv = tf.nn.conv2d(input, weights,
strides=[1, 1, 1, 1], padding='SAME')
return tf.nn.relu(conv + biases)
# 定义一个图片过滤器
def my_image_filter(input_images):
with tf.variable_scope("conv1"):
# Variables created here will be named "conv1/weights", "conv1/biases".
relu1 = conv_relu(input_images, [5, 5, 32, 32], [32])
with tf.variable_scope("conv2"):
# Variables created here will be named "conv2/weights", "conv2/biases".
return conv_relu(relu1, [5, 5, 32, 32], [32])
# 实验一:调用 my_image_filter() 两次
result1 = my_image_filter(image1)
result2 = my_image_filter(image2)
>>> Raises ValueError(... conv1/weights already exists ...), tf.get_variable()会检测已经存在的变量是否已经共享
# 解决方法一, 可以在设计网络时加上一个布尔型的 reuse 参数
with tf.variable_scope("image_filters"):
result1 = my_image_filter(image1)
with tf.variable_scope("image_filters", reuse=True):
result2 = my_image_filter(image2)
# 解决方法二
with tf.variable_scope("image_filters") as scope:
# 下面我们两次调用 my_image_filter 函数,但是由于引入了变量共享机制
# 可以看到我们只是创建了一遍网络结构。
result1 = my_image_filter(image1)
scope.reuse_variables()
result2 = my_image_filter(image2)
# 解决方法三
with tf.variable_scope("image_filters") as scope:
result1 = my_image_filter(image1)
with tf.variable_scope(scope, reuse=True):
result2 = my_image_filter(image2)
# 打印出所有的可训练参变量
vs = tf.trainable_variables()
print('There are %d trainable_variables in the Graph: ' % len(vs))
for v in vs:
print(v)
# 输出结果证明确实:参变量共享,因为只有四个变量,没有创建新的变量。
There are 4 trainable_variables in the Graph:
Tensor("image_filters/conv1/weights/read:0", shape=(5, 5, 32, 32), dtype=float32)
Tensor("image_filters/conv1/biases/read:0", shape=(32,), dtype=float32)
Tensor("image_filters/conv2/weights/read:0", shape=(5, 5, 32, 32), dtype=float32)
Tensor("image_filters/conv2/biases/read:0", shape=(32,), dtype=float32)
# Returns all variables created with trainable=True in a var_list
var_list = tf.trainable_variables()
init = tf.global_variables_initializer()
sess.run(init)
for var in var_list:
sess.run(var)
1、https://www.tensorflow.org/programmers_guide/variables#sharing_variables
2、共享变量
3、TensorFlow 变量与命名空间_固有结界