Tensorflow替换静态图中的OP

浅浅的花香味﹌ 2021-11-01 10:26 304阅读 0赞

import tensorflow as tf

import collections

from tensorflow.core.framework import tensor\_shape\_pb2

\# 读取模型

graph\_def = tf.GraphDef()

with tf.gfile.FastGFile('./pb/model.pb', 'rb') as f:

graph\_def.ParseFromString(f.read())

\# 统计图中的node，保存为map.其中 key : value = op.name : op

input\_node\_map = \{\}

for node in graph\_def.node:

if node.name not in input\_node\_map.keys():

input\_node\_map\[node.name\] = node

else:

raise ValueError("Duplicate node names detected for ", node.name)

\# 统计每一个op被使用的次数

node\_reference\_count = collections.defaultdict(int)

output\_node\_names = \['xnet/Softmax'\]

for node in graph\_def.node:

for input\_name in node.input:

stripped\_name = input\_name

node\_reference\_count\[stripped\_name\] += 1

for output\_name in output\_node\_names:

node\_reference\_count\[output\_name\] += 1

\# 删除old\_op

old\_op = input\_node\_map\['xnet/Layer\_Conv\_1/Conv2D'\]

node\_reference\_count\['xnet/Layer\_Conv\_1/Conv2D'\] -= 1

\# 创建新的op

new\_node = tf.NodeDef()

new\_node.op = 'Conv2D'

new\_node.name = 'new\_Conv\_1'

for input\_name in old\_op.input:

new\_node.input.extend(\[input\_name\])

new\_node.attr\["T"\].CopyFrom(tf.AttrValue(type=tf.float32.as\_datatype\_enum)) \# (old\_op.attr\["T"\])

new\_node.attr\["use\_cudnn\_on\_gpu"\].CopyFrom(tf.AttrValue(b=1)) \# (old\_op.attr\["use\_cudnn\_on\_gpu"\])

stride\_list = \[1, 2, 2, 1\]

new\_node.attr\["strides"\].CopyFrom(tf.AttrValue(list=tf.AttrValue.ListValue(i=stride\_list))) \# (old\_op.attr\["strides"\])

new\_node.attr\["padding"\].CopyFrom(tf.AttrValue(s=b'VALID')) \# (old\_op.attr\["padding"\])

\# 创建const类型的op，仅作为测试，本实验中不添加入graph

new\_const = tf.NodeDef()

new\_const.op = 'Const'

new\_const.name = 'new\_Const'

new\_const.attr\['dtype'\].CopyFrom(tf.AttrValue(type=tf.float32.as\_datatype\_enum))

new\_const.attr\['value'\].CopyFrom(

tf.AttrValue(tensor=tf.make\_tensor\_proto(\[4, 5, 0, 0, 8, 0, 7, 0\], tf.float32, \[4, 2\])))

new\_const.attr\['\_output\_shapes'\].CopyFrom(

tf.AttrValue(list=tf.AttrValue.ListValue(shape=\[tensor\_shape\_pb2.TensorShapeProto(

dim=\[tensor\_shape\_pb2.TensorShapeProto.Dim(size=4), tensor\_shape\_pb2.TensorShapeProto.Dim(size=2)\])\])))

\# 将new\_node作为输入赋值给图中节点

for node in graph\_def.node:

if old\_op.name in node.input:

for i, name in enumerate(node.input):

if name == old\_op.name:

node.input\[i\] = new\_node.name

print('success\_1')

\# 定义一个新图

graph\_def\_new = tf.GraphDef()

for node in graph\_def.node:

if node\_reference\_count\[node.name\] < 1:

continue

new = tf.NodeDef()

new.CopyFrom(node)

graph\_def\_new.node.extend(\[new\])

graph\_def\_new.node.extend(\[new\_node\])

\# graph\_def\_new.node.extend(\[new\_const\])

\# 将新图注入到默认的Graph中

tf.import\_graph\_def(graph\_def\_new, name='') \# Imports \`graph\_def\` into the current default \`Graph\`

\# 测试案例

with tf.Session() as sess:

tf.train.write\_graph(sess.graph\_def, logdir='./pb', name='graph\_def\_new.pb')

OP的信息：

name: "xnet/Layer\_FC\_32/xw\_plus\_b"

op: "BiasAdd"

input: "xnet/Layer\_FC\_32/xw\_plus\_b/MatMul"

input: "xnet/Layer\_FC\_32/biases/read"

attr

key: "T"

value \{type: DT\_FLOAT\}

attr

key: "data\_format"

value \{s: "NHWC"\}

在tensorflow中，OP主要包括以下信息：name, op , input, attr

name--类型string。 在模型定义的时候由工程师定义，如果工程师没有定义的话会自动的利用op作为其值

op--类型string。表示这是一个什么op,比如加减乘除，当在运行的时候，编译器会更具op调用相应的算子来做计算

input--类型list.。列表中包含了该节点输入，是有序的，不可以被assign

attr--类型map。map中的key和value一般是指该OP的配置信息

OP的操作：

1、op信息获取

1. 通过Graph获取op

op = tf.get\_default\_graph().get\_Operations()

print(op\[0\])

print(op\[0\].name)

\# 如果想获得属性或者input信息需要如下写法

print(op\[0\].node\_def.attr)

2.通过Graph\_def获取op

op = graph\_def.node

print(op\[0\].name)

print(op\[0\].input)

2、op的创建无锡好的男科医院 http://www.zzchnk.com/

在构建新的op的时候需要对op的属性比较清楚，对于没有default的属性一定要做好初始化

1.根据已有的op创建新的op

new\_node = tf.NodeDef() \# 构建一个op对象，所有属性都为空

new\_node.op = 'Conv2D'

new\_node.name = 'new\_Conv\_1'

for input\_name in old\_op.input: \# 原始op的input导入进来

new\_node.input.extend(\[input\_name\])

new\_node.attr\["T"\].CopyFrom(old\_op.attr\["T"\])

new\_node.attr\["use\_cudnn\_on\_gpu"\].CopyFrom(old\_op.attr\["use\_cudnn\_on\_gpu"\])

new\_node.attr\["strides"\].CopyFrom(old\_op.attr\["strides"\])

new\_node.attr\["padding"\].CopyFrom(old\_op.attr\["padding"\])

2.创建一个自定义的op

new\_op = tf.NodeDef()

new\_op.op = "Const"

new\_op.name = conv\_op.name

new\_op.attr\["dtype"\].CopyFrom(tf.AttrValue( type=tf.int32.as\_datatype\_enum))

new\_op.attr\["value"\].CopyFrom(tf.AttrValue(tensor=tf.make\_tensor\_proto(\[0, 0, 0, 0, 0, 0, 0, 0\], tf.int32, \[4, 2\])))

OP中attr为map,每一个map中key为字符串，value为的类型由下面9种，每种对应的原型如下表所示：

repeated bytes s = 2; // "list(string)"

repeated int64 i = 3 \[packed = true\]; // "list(int)"

repeated float f = 4 \[packed = true\]; // "list(float)"

repeated bool b = 5 \[packed = true\]; // "list(bool)"

repeated DataType type = 6 \[packed = true\]; // "list(type)"

repeated TensorShapeProto shape = 7; // "list(shape)"

repeated TensorProto tensor = 8; // "list(tensor)"

repeated NameAttrList func = 9; // "list(attr)"

list 也为value的一种类型

每一种类型初始化方式：

CopyFrom(tf.AttrValue( s=b'hello,world'))

CopyFrom(tf.AttrValue( i=88 ))

CopyFrom(tf.AttrValue( f=88.0 ))

CopyFrom(tf.AttrValue( b=1/0 ))

new\_op.attr\["dtype"\].CopyFrom(tf.AttrValue( type=tf.int32.as\_datatype\_enum))

from tensorflow.core.framework import tensor\_shape\_pb2

tensor\_shape\_pb2.TensorShapeProto(dim=\[tensor\_shape\_pb2.TensorShapeProto.Dim(size=-1 if d.value is None else d.value) for d in dims\])

new\_op.attr\["value"\].CopyFrom(tf.AttrValue(tensor=tf.make\_tensor\_proto(\[0, 0, 0, 0, 0, 0, 0, 0\], tf.int32, \[4, 2\])))

func目前没有没有遇到过

stride\_list = \[1, 2, 2, 1\]

new\_node.attr\["strides"\].CopyFrom(tf.AttrValue(list=tf.AttrValue.ListValue(i=stride\_list)))

import tensorflow as tf

import collections

from tensorflow.core.framework import tensor\_shape\_pb2

\# 读取模型

graph\_def = tf.GraphDef()

with tf.gfile.FastGFile('./pb/model.pb', 'rb') as f:

graph\_def.ParseFromString(f.read())

\# 统计图中的node，保存为map.其中 key : value = op.name : op

input\_node\_map = \{\}

for node in graph\_def.node:

if node.name not in input\_node\_map.keys():

input\_node\_map\[node.name\] = node

else:

raise ValueError("Duplicate node names detected for ", node.name)

\# 统计每一个op被使用的次数

node\_reference\_count = collections.defaultdict(int)

output\_node\_names = \['xnet/Softmax'\]

for node in graph\_def.node:

for input\_name in node.input:

stripped\_name = input\_name

node\_reference\_count\[stripped\_name\] += 1

for output\_name in output\_node\_names:

node\_reference\_count\[output\_name\] += 1

\# 删除old\_op

old\_op = input\_node\_map\['xnet/Layer\_Conv\_1/Conv2D'\]

node\_reference\_count\['xnet/Layer\_Conv\_1/Conv2D'\] -= 1

\# 创建新的op

new\_node = tf.NodeDef()

new\_node.op = 'Conv2D'

new\_node.name = 'new\_Conv\_1'

for input\_name in old\_op.input:

new\_node.input.extend(\[input\_name\])

new\_node.attr\["T"\].CopyFrom(tf.AttrValue(type=tf.float32.as\_datatype\_enum)) \# (old\_op.attr\["T"\])

new\_node.attr\["use\_cudnn\_on\_gpu"\].CopyFrom(tf.AttrValue(b=1)) \# (old\_op.attr\["use\_cudnn\_on\_gpu"\])

stride\_list = \[1, 2, 2, 1\]

new\_node.attr\["strides"\].CopyFrom(tf.AttrValue(list=tf.AttrValue.ListValue(i=stride\_list))) \# (old\_op.attr\["strides"\])

new\_node.attr\["padding"\].CopyFrom(tf.AttrValue(s=b'VALID')) \# (old\_op.attr\["padding"\])

\# 创建const类型的op，仅作为测试，本实验中不添加入graph

new\_const = tf.NodeDef()

new\_const.op = 'Const'

new\_const.name = 'new\_Const'

new\_const.attr\['dtype'\].CopyFrom(tf.AttrValue(type=tf.float32.as\_datatype\_enum))

new\_const.attr\['value'\].CopyFrom(

tf.AttrValue(tensor=tf.make\_tensor\_proto(\[4, 5, 0, 0, 8, 0, 7, 0\], tf.float32, \[4, 2\])))

new\_const.attr\['\_output\_shapes'\].CopyFrom(

tf.AttrValue(list=tf.AttrValue.ListValue(shape=\[tensor\_shape\_pb2.TensorShapeProto(

dim=\[tensor\_shape\_pb2.TensorShapeProto.Dim(size=4), tensor\_shape\_pb2.TensorShapeProto.Dim(size=2)\])\])))

\# 将new\_node作为输入赋值给图中节点

for node in graph\_def.node:

if old\_op.name in node.input:

for i, name in enumerate(node.input):

if name == old\_op.name:

node.input\[i\] = new\_node.name

print('success\_1')

\# 定义一个新图

graph\_def\_new = tf.GraphDef()

for node in graph\_def.node:

if node\_reference\_count\[node.name\] < 1:

continue

new = tf.NodeDef()

new.CopyFrom(node)

graph\_def\_new.node.extend(\[new\])

graph\_def\_new.node.extend(\[new\_node\])

\# graph\_def\_new.node.extend(\[new\_const\])

\# 将新图注入到默认的Graph中

tf.import\_graph\_def(graph\_def\_new, name='') \# Imports \`graph\_def\` into the current default \`Graph\`

\# 测试案例

with tf.Session() as sess:

tf.train.write\_graph(sess.graph\_def, logdir='./pb', name='graph\_def\_new.pb')

OP的信息：

name: "xnet/Layer\_FC\_32/xw\_plus\_b"

op: "BiasAdd"

input: "xnet/Layer\_FC\_32/xw\_plus\_b/MatMul"

input: "xnet/Layer\_FC\_32/biases/read"

attr

key: "T"

value \{type: DT\_FLOAT\}

attr

key: "data\_format"

value \{s: "NHWC"\}

在tensorflow中，OP主要包括以下信息：name, op , input, attr

name--类型string。 在模型定义的时候由工程师定义，如果工程师没有定义的话会自动的利用op作为其值

op--类型string。表示这是一个什么op,比如加减乘除，当在运行的时候，编译器会更具op调用相应的算子来做计算

input--类型list.。列表中包含了该节点输入，是有序的，不可以被assign

attr--类型map。map中的key和value一般是指该OP的配置信息

OP的操作：

1、op信息获取

1. 通过Graph获取op

op = tf.get\_default\_graph().get\_Operations()

print(op\[0\])

print(op\[0\].name)

\# 如果想获得属性或者input信息需要如下写法

print(op\[0\].node\_def.attr)

2.通过Graph\_def获取op

op = graph\_def.node

print(op\[0\].name)

print(op\[0\].input)

2、op的创建

在构建新的op的时候需要对op的属性比较清楚，对于没有default的属性一定要做好初始化

1.根据已有的op创建新的op

new\_node = tf.NodeDef() \# 构建一个op对象，所有属性都为空

new\_node.op = 'Conv2D'

new\_node.name = 'new\_Conv\_1'

for input\_name in old\_op.input: \# 原始op的input导入进来

new\_node.input.extend(\[input\_name\])

new\_node.attr\["T"\].CopyFrom(old\_op.attr\["T"\])

new\_node.attr\["use\_cudnn\_on\_gpu"\].CopyFrom(old\_op.attr\["use\_cudnn\_on\_gpu"\])

new\_node.attr\["strides"\].CopyFrom(old\_op.attr\["strides"\])

new\_node.attr\["padding"\].CopyFrom(old\_op.attr\["padding"\])

2.创建一个自定义的op

new\_op = tf.NodeDef()

new\_op.op = "Const"

new\_op.name = conv\_op.name

new\_op.attr\["dtype"\].CopyFrom(tf.AttrValue( type=tf.int32.as\_datatype\_enum))

new\_op.attr\["value"\].CopyFrom(tf.AttrValue(tensor=tf.make\_tensor\_proto(\[0, 0, 0, 0, 0, 0, 0, 0\], tf.int32, \[4, 2\])))

OP中attr为map,每一个map中key为字符串，value为的类型由下面9种，每种对应的原型如下表所示：

repeated bytes s = 2; // "list(string)"

repeated int64 i = 3 \[packed = true\]; // "list(int)"

repeated float f = 4 \[packed = true\]; // "list(float)"

repeated bool b = 5 \[packed = true\]; // "list(bool)"

repeated DataType type = 6 \[packed = true\]; // "list(type)"

repeated TensorShapeProto shape = 7; // "list(shape)"

repeated TensorProto tensor = 8; // "list(tensor)"

repeated NameAttrList func = 9; // "list(attr)"

list 也为value的一种类型

每一种类型初始化方式：

CopyFrom(tf.AttrValue( s=b'hello,world'))

CopyFrom(tf.AttrValue( i=88 ))

CopyFrom(tf.AttrValue( f=88.0 ))

CopyFrom(tf.AttrValue( b=1/0 ))

new\_op.attr\["dtype"\].CopyFrom(tf.AttrValue( type=tf.int32.as\_datatype\_enum))

from tensorflow.core.framework import tensor\_shape\_pb2

tensor\_shape\_pb2.TensorShapeProto(dim=\[tensor\_shape\_pb2.TensorShapeProto.Dim(size=-1 if d.value is None else d.value) for d in dims\])

new\_op.attr\["value"\].CopyFrom(tf.AttrValue(tensor=tf.make\_tensor\_proto(\[0, 0, 0, 0, 0, 0, 0, 0\], tf.int32, \[4, 2\])))

func目前没有没有遇到过

stride\_list = \[1, 2, 2, 1\]

new\_node.attr\["strides"\].CopyFrom(tf.AttrValue(list=tf.AttrValue.ListValue(i=stride\_list)))

转载于:https://www.cnblogs.com/djw12333/p/11268947.html

Tensorflow替换静态图中的OP

发表评论取消回复

还没有评论，来说两句吧...

相关阅读

相关 tensorflow的用法：data_flow_ops.FIFOQueue的用法

相关 tensorflow:OP_REQUIRES failed at save_restore_v2_ops.cc:109

相关 tensorflow c++ op 生成 python调用接口

相关在TensorFlow2.x中执行TensorFlow1.x代码的静态图执行模式

相关【tensorflow】自定义op

相关 Python TensorFlow，计算图，创建计算图，Graph，运算操作(op)

相关 tensorflow中control_flow_ops.while_loop

相关 tensorflow中control_flow_ops.while_loop

相关 Tensorflow替换静态图中的OP

相关 tensorflow 在优化器后面添加bn的ops依赖