Question

我有一个frozen inference graph存储在.pb file中，该trained Tensorflow model是freeze_graph函数从sigmoid activations获得的。

假设，为简单起见，我想将模型中的部分tanh activations更改为Graph Editor library in tf.contrib（并且不要讨论是否这是一个好主意）。

如何才能只访问.pb文件中的冻结图表，并且无法重新训练模型？

我知道<ui:composition xmlns="http://www.w3.org/1999/xhtml" xmlns:h="http://java.sun.com/jsf/html" xmlns:f="http://java.sun.com/jsf/core" xmlns:ui="http://java.sun.com/jsf/facelets" xmlns:p="http://primefaces.org/ui"> <header id="header" class="hoc clear">  <nav id="mainav" class="fl_left"> <ul class="clear"> <li class="#{funksionsBean.getIndex()}"><a href="index.xhtml">Faqja Kryesore</a></li> <li class="#{funksionsBean.getDestinations()}"><a href="destinations.xhtml">Destinacionet Turistike</a></li> <li class="#{funksionsBean.getContact()}"><a href="contact.xhtml">Na kontaktoni</a></li> </ul> </nav> <div id="logo" class="fl_right"> <h:form id="accountForm"> <h:panelGrid columns="2" style="background: transparent" rendered="#{!clientBean.logged}"> <p:commandButton value="HYR" styleClass="login" onclick="PF('logInDialog').show();" /> <p:commandButton value="REGJISTROHU" styleClass="login" /> </h:panelGrid> <h:panelGrid columns="3" style="background: transparent" rendered="#{clientBean.logged}"> <h:outputLabel value="PERSHENDETJE, #{clientBean.client.getClient_name()}" styleClass="pershendetje"></h:outputLabel> <p:commandButton value="Llogaria" styleClass="login" action="llogaria.xhtml?faces-redirect=true" /> <p:commandButton value="DIL" styleClass="login" action="#{clientBean.LogOut}" /> </h:panelGrid> <p:dialog header="Ju lutem fusni te dhenat" widgetVar="logInDialog" minHeight="40"> <h:panelGrid id="inputPanel" columns="2" cellpadding="10" styleClass="logInPannel"> <h:outputLabel value="Email:" /> <h:inputText value="#{clientBean.email}" /> <h:outputLabel value="Password:" /> <h:inputSecret value="#{clientBean.password}"></h:inputSecret> </h:panelGrid> <h:commandButton action="#{clientBean.LogIn}" value="Hyr"></h:commandButton> </p:dialog> </h:form> </div> </header>应该可以做这种工作，但我无法在文档中找到一种简单的方法。

Answer 1

解决方案是使用import_graph_def：

import tensorflow as tf
sess = tf.Session()

def load_graph(frozen_graph_filename):
    with tf.gfile.GFile(frozen_graph_filename, "rb") as f:
        graph_def = tf.GraphDef()
        graph_def.ParseFromString(f.read())
    with tf.Graph().as_default() as graph:
        tf.import_graph_def(graph_def, name='')
    return graph
graph_model = load_graph("frozen_inference_graph.pb")
graph_model_def = graph_model.as_graph_def()

graph_new = tf.Graph()
graph_new.as_default()
my_new_tensor = # whatever
tf.import_graph_def(graph_model_def, name='', input_map={"tensor_to_replace": my_new_tensor})
#do somthing with your new graph

Here我写了一篇关于它的帖子

Answer 2

* .pb文件包含SavedModel协议缓冲区。您应该可以使用SavedModel loader加载它。您也可以使用SavedModel CLI进行操作。 SavedModels的完整文档是here。

Answer 3

你可以试试这个：

graph = load_graph(filename)
graph_def = graph.as_graph_def()
# if ReLu op is at node 161
graph_def.node[161].op="tanh"
tf.train.write_graph(graph_def, path2savfrozn, "altered_frozen.pb", False)

请知道它是否有效。

Answer 4

遵循这些原则应该可以起作用：

graph_def = tf.GraphDef()
with open('frozen_inference.pb', 'rb') as f:
    graph_def.ParseFromString(f.read())

with tf.Graph().as_default() as graph:
    importer.import_graph_def(graph_def, name='')

new_model = tf.GraphDef()

with tf.Session(graph=graph) as sess:

    for n in sess.graph_def.node:

        if n.op == 'Sigmoid':
            nn = new_model.node.add()
            nn.op = 'Tanh'
            nn.name = n.name
            for i in n.input:
                nn.input.extend([i])

        else:
            nn = new_model.node.add()
            nn.CopyFrom(n)

替换冻结的Tensorflow模型中的节点

4 个答案: