tensorflow graph 中获取variable operation

tensorflow提供了一些列的方法获取和变量计算图中的variable和operation。

获取单个operation/variable

可以通过如下两个方法获取图中的相关variable和operation：

1. tf.Graph.get_tensor_by_name(tensor_name)

2. tf.Graph.get_operation_by_name(op_name)

批量获取

批量获取的方式主要有如下几种：

1. graph.node

import tensorflow as tf

# Create some variables.
v1 = tf.get_variable("v1", shape=[3], initializer = tf.zeros_initializer)
v2 = tf.get_variable("v2", shape=[5], initializer = tf.zeros_initializer)
v3 = tf.get_variable("v3", shape=[4], initializer = tf.zeros_initializer)

inc_v1 = tf.assign(v1,v1+1,name='inc_v1')
dec_v2 = tf.assign(v2,v2-1,name='dec_v2')
dec_v3 = tf.assign(v3,v3-2,name='dec_v3')

# Add an op to initialize the variables.
init_op = tf.global_variables_initializer()

# Add ops to save and restore all the variables.
saver = tf.train.Saver()

# Later, launch the model, initialize the variables, do some work, and save the
# variables to disk.
with tf.Session() as sess:

sess.run(init_op)
# Do some work with the model.
inc_v1.op.run()
dec_v2.op.run()
dec_v3.op.run()

for n in tf.get_default_graph().as_graph_def().node:
print n

输出：

name: "v1/Initializer/zeros"
op: "Const"
attr {
key: "_class"
value {
list {
s: "loc:@v1"
}
}
}
attr {
key: "dtype"
value {
type: DT_FLOAT
}
}
attr {
key: "value"
value {
tensor {
dtype: DT_FLOAT
tensor_shape {
dim {
size: 3
}
}
float_val: 0.0
}
}
}

该方法列出了每个graph中每个node的详细信息。

2. graph.get_operations()

for op in tf.get_default_graph().get_operations():
print op.name
print op.values()

输出：

name:v1/Initializer/zeros
value:(<tf.Tensor 'v1/Initializer/zeros:0' shape=(3,) dtype=float32>,)
name:v1
value:(<tf.Tensor 'v1:0' shape=(3,) dtype=float32_ref>,)

op.valuses()将返回该op对应的tensor对象，可以进一步获取tensor的name，shape等信息。

3. tf.all_variables()

import tensorflow as tf

# Create some variables.
v1 = tf.get_variable("v1", shape=[3], initializer = tf.zeros_initializer)
v2 = tf.get_variable("v2", shape=[5], initializer = tf.zeros_initializer)
v3 = tf.get_variable("v3", shape=[4], initializer = tf.zeros_initializer)

inc_v1 = tf.assign(v1,v1+1,name='inc_v1')
dec_v2 = tf.assign(v2,v2-1,name='dec_v2')
dec_v3 = tf.assign(v3,v3-2,name='dec_v3')

# Add an op to initialize the variables.
init_op = tf.global_variables_initializer()

# Add ops to save and restore all the variables.
saver = tf.train.Saver()

# Later, launch the model, initialize the variables, do some work, and save the
# variables to disk.
with tf.Session() as sess:

sess.run(init_op)
# Do some work with the model.
inc_v1.op.run()
dec_v2.op.run()
dec_v3.op.run()

for variable in tf.all_variables():
print variable
print variable.name

输出：

<tf.Variable 'v1:0' shape=(3,) dtype=float32_ref>
v1:0
<tf.Variable 'v2:0' shape=(5,) dtype=float32_ref>
v2:0
<tf.Variable 'v3:0' shape=(4,) dtype=float32_ref>
v3:0

该方法返回默认计算图中所有的variable（）对象

4. tf.get_collection(collection_key)

for variable in tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES):
print variable

输出：

<tf.Variable 'v1:0' shape=(3,) dtype=float32_ref>
<tf.Variable 'v2:0' shape=(5,) dtype=float32_ref>
<tf.Variable 'v3:0' shape=(4,) dtype=float32_ref>

该方法根据key返回相应collection中的对象。

tensorflow中预定义了一些grapykClass GraphKeys

Standard names to use for graph collections.
The standard library uses various well-known names to collect and retrieve values associated with a graph. For example, the tf.Optimizer subclasses default to optimizing the variables collected under tf.GraphKeys.TRAINABLE_VARIABLES if none is specified, but it is also possible to pass an explicit list of variables.
The following standard keys are defined:
● GLOBAL_VARIABLES: the default collection of Variable objects, shared across distributed environment (model variables are subset of these). See tf.global_variables for more details. Commonly, all TRAINABLE_VARIABLES variables will be in MODEL_VARIABLES, and all MODEL_VARIABLES variables will be in GLOBAL_VARIABLES.
● LOCAL_VARIABLES: the subset of Variable objects that are local to each machine. Usually used for temporarily variables, like counters. Note: use tf.contrib.framework.local_variable to add to this collection.
● MODEL_VARIABLES: the subset of Variable objects that are used in the model for inference (feed forward). Note: use tf.contrib.framework.model_variable to add to this collection.
● TRAINABLE_VARIABLES: the subset of Variable objects that will be trained by an optimizer. Seetf.trainable_variables for more details.
● SUMMARIES: the summary Tensor objects that have been created in the graph. See tf.summary.merge_all for more details.
● QUEUE_RUNNERS: the QueueRunner objects that are used to produce input for a computation. Seetf.train.start_queue_runners for more details.
● MOVING_AVERAGE_VARIABLES: the subset of Variable objects that will also keep moving averages. Seetf.moving_average_variables for more details.
● REGULARIZATION_LOSSES: regularization losses collected during graph construction.
The following standard keys are defined, but their collections are not automatically populated as many of the others are:
● WEIGHTS
● BIASES
● ACTIVATIONS

除了预定义的collecion，tensorflow还支持自定义collection方法–tf.add_collection(key,value),tf.get_collection(key)。tf的collecion提供了一种全局的存储机制，不收命名空间影响。代码如下：

import tensorflow as tf

# Create some variables.
v1 = tf.get_variable("v1", shape=[3], initializer = tf.zeros_initializer)
v2 = tf.get_variable("v2", shape=[5], initializer = tf.zeros_initializer)
v3 = tf.get_variable("v3", shape=[4], initializer = tf.zeros_initializer)

inc_v1 = tf.assign(v1,v1+1,name='inc_v1')
dec_v2 = tf.assign(v2,v2-1,name='dec_v2')
dec_v3 = tf.assign(v3,v3-2,name='dec_v3')

# Add an op to initialize the variables.
init_op = tf.global_variables_initializer()

# Add ops to save and restore all the variables.
saver = tf.train.Saver()

# Later, launch the model, initialize the variables, do some work, and save the
# variables to disk.
with tf.Session() as sess:

sess.run(init_op)
# Do some work with the model.
inc_v1.op.run()
dec_v2.op.run()
dec_v3.op.run()

# Add variable into
tf.add_to_collection('test',v1)
tf.add_to_collection('test',v2)
tf.add_to_collection('test',inc_v1)

for element in tf.get_collection('test'):
print element

输出：

<tf.Variable 'v1:0' shape=(3,) dtype=float32_ref>
<tf.Variable 'v2:0' shape=(5,) dtype=float32_ref>
Tensor("inc_v1:0", shape=(3,), dtype=float32_ref)

tf还提供了获取graph中所有collection的方法：

for key in tf.get_default_graph().get_all_collection_keys():
print 'key:'+key
for element in tf.get_collection(key):
print element

输出

   key:variables
<tf.Variable 'v1:0' shape=(3,) dtype=float32_ref>
<tf.Variable 'v2:0' shape=(5,) dtype=float32_ref>
<tf.Variable 'v3:0' shape=(4,) dtype=float32_ref>
key:trainable_variables
<tf.Variable 'v1:0' shape=(3,) dtype=float32_ref>
<tf.Variable 'v2:0' shape=(5,) dtype=float32_ref>
<tf.Variable 'v3:0' shape=(4,) dtype=float32_ref>