blocks|key|4350763|text|TensorFlow本身并不包含任何将TensorFlow图形导出为协议缓冲区文件的方法，但您可以使用常规的Keras实用程序来实现这一点。Here是一篇博客文章，解释了如何使用TensorFlow中包含的实用程序脚本freeze_graph.py来完成它，这是它的“典型”完成方式。|type|unstyled|depth|inlineStyleRanges|offset|length|style|CODE|entityRanges|data|4350764|然而，我个人觉得创建一个检查点，然后运行外部脚本来获取模型是一件麻烦的事情，而不是更喜欢从我自己的Python代码中完成，所以我使用这样的函数：|4350765|def+freeze_session(session,+keep_var_names=None,+output_names=None,+clear_devices=True):
++++"""
++++Freezes+the+state+of+a+session+into+a+pruned+computation+graph.

++++Creates+a+new+computation+graph+where+variable+nodes+are+replaced+by
++++constants+taking+their+current+value+in+the+session.+The+new+graph+will+be
++++pruned+so+subgraphs+that+are+not+necessary+to+compute+the+requested
++++outputs+are+removed.
++++@param+session+The+TensorFlow+session+to+be+frozen.
++++@param+keep_var_names+A+list+of+variable+names+that+should+not+be+frozen,
++++++++++++++++++++++++++or+None+to+freeze+all+the+variables+in+the+graph.
++++@param+output_names+Names+of+the+relevant+graph+outputs.
++++@param+clear_devices+Remove+the+device+directives+from+the+graph+for+better+portability.
++++@return+The+frozen+graph+definition.
++++"""
++++graph+=+session.graph
++++with+graph.as_default():
++++++++freeze_var_names+=+list(set(v.op.name+for+v+in+tf.global_variables()).difference(keep_var_names+or+[]))
++++++++output_names+=+output_names+or+[]
++++++++output_names+%2B=+[v.op.name+for+v+in+tf.global_variables()]
++++++++input_graph_def+=+graph.as_graph_def()
++++++++if+clear_devices:
++++++++++++for+node+in+input_graph_def.node:
++++++++++++++++node.device+=+""
++++++++frozen_graph+=+tf.graph_util.convert_variables_to_constants(
++++++++++++session,+input_graph_def,+output_names,+freeze_var_names)
++++++++return+frozen_graph|code-block|syntax|javascript|4350766|这是在freeze_graph.py的实现中受到启发的。参数也类似于脚本。session是TensorFlow会话对象。只有当你想保持某些变量不被冻结(例如，对于有状态模型)时，才需要keep_var_names，所以通常不需要。output_names是一个列表，其中包含生成所需输出的操作的名称。clear_devices只是删除了任何设备指令，以使图形更易于移植。因此，对于具有一个输出的典型Keras+model，您将执行如下操作：|4350767|from+keras+import+backend+as+K

#+Create,+compile+and+train+model...

frozen_graph+=+freeze_session(K.get_session(),
++++++++++++++++++++++++++++++output_names=[out.op.name+for+out+in+model.outputs])|4350768|然后，您可以像往常一样使用tf.train.write_graph将图形写入文件|4350769|tf.train.write_graph(frozen_graph,+"some_directory",+"my_model.pb",+as_text=False)|4350770|entityMap|0|LINK|mutability|MUTABLE|url|https://blog.metaflow.fr/tensorflow-how-to-freeze-a-model-and-serve-it-with-a-python-api-d4f3596b3adc|1|https://github.com/tensorflow/tensorflow/blob/v1.12.0/tensorflow/python/tools/freeze_graph.py|2|https://www.tensorflow.org/api_docs/python/tf/train/write_graph^0|31|F|1Y|4|0|31|F|1|0|0|0|3|F|11|7|2K|E|37|C|46|D|5P|5|0|0|D|K|D|K|2|0|0^^$0|@$1|2|3|4|5|6|7|16|8|@$9|17|A|18|B|C]]|D|@$9|19|A|1A|1|1B]|$9|1C|A|1D|1|1E]]|E|$]]|$1|F|3|G|5|6|7|1F|8|@]|D|@]|E|$]]|$1|H|3|I|5|J|7|1G|8|@]|D|@]|E|$K|L]]|$1|M|3|N|5|6|7|1H|8|@$9|1I|A|1J|B|C]|$9|1K|A|1L|B|C]|$9|1M|A|1N|B|C]|$9|1O|A|1P|B|C]|$9|1Q|A|1R|B|C]|$9|1S|A|1T|B|C]]|D|@]|E|$]]|$1|O|3|P|5|J|7|1U|8|@]|D|@]|E|$K|L]]|$1|Q|3|R|5|6|7|1V|8|@$9|1W|A|1X|B|C]]|D|@$9|1Y|A|1Z|1|20]]|E|$]]|$1|S|3|T|5|J|7|21|8|@]|D|@]|E|$K|L]]|$1|U|3|-4|5|6|7|22|8|@]|D|@]|E|$]]]|V|$W|$5|X|Y|Z|E|$10|11]]|12|$5|X|Y|Z|E|$10|13]]|14|$5|X|Y|Z|E|$10|15]]]]

Keras does not include by itself any means to export a TensorFlow graph as a protocol buffers file, but you can do it using regular TensorFlow utilities. <a href="https://blog.metaflow.fr/tensorflow-how-to-freeze-a-model-and-serve-it-with-a-python-api-d4f3596b3adc" rel="noreferrer">Here</a> is a blog post explaining how to do it using the utility script <a href="https://github.com/tensorflow/tensorflow/blob/v1.12.0/tensorflow/python/tools/freeze_graph.py" rel="noreferrer"><code>freeze_graph.py</code></a> included in TensorFlow, which is the "typical" way it is done.

However, I personally find a nuisance having to make a checkpoint and then run an external script to obtain a model, and instead prefer to do it from my own Python code, so I use a function like this:

<pre class="lang-py prettyprint-override"><code>def freeze_session(session, keep_var_names=None, output_names=None, clear_devices=True):
 """
 Freezes the state of a session into a pruned computation graph.

 Creates a new computation graph where variable nodes are replaced by
 constants taking their current value in the session. The new graph will be
 pruned so subgraphs that are not necessary to compute the requested
 outputs are removed.
 @param session The TensorFlow session to be frozen.
 @param keep_var_names A list of variable names that should not be frozen,
 or None to freeze all the variables in the graph.
 @param output_names Names of the relevant graph outputs.
 @param clear_devices Remove the device directives from the graph for better portability.
 @return The frozen graph definition.
 """
 graph = session.graph
 with graph.as_default():
 freeze_var_names = list(set(v.op.name for v in tf.global_variables()).difference(keep_var_names or []))
 output_names = output_names or []
 output_names += [v.op.name for v in tf.global_variables()]
 input_graph_def = graph.as_graph_def()
 if clear_devices:
 for node in input_graph_def.node:
 node.device = ""
 frozen_graph = tf.graph_util.convert_variables_to_constants(
 session, input_graph_def, output_names, freeze_var_names)
 return frozen_graph
</code></pre>

Which is inspired in the implementation of <code>freeze_graph.py</code>. The parameters are similar to the script too. <code>session</code> is the TensorFlow session object. <code>keep_var_names</code> is only needed if you want to keep some variable not frozen (e.g. for stateful models), so generally not. <code>output_names</code> is a list with the names of the operations that produce the outputs that you want. <code>clear_devices</code> just removes any device directives to make the graph more portable. So, for a typical Keras <code>model</code> with one output, you would do something like:

<pre><code>from keras import backend as K

# Create, compile and train model...

frozen_graph = freeze_session(K.get_session(),
 output_names=[out.op.name for out in model.outputs])
</code></pre>

Then you can write the graph to a file as usual with <a href="https://www.tensorflow.org/api_docs/python/tf/train/write_graph" rel="noreferrer"><code>tf.train.write_graph</code></a>:

<pre><code>tf.train.write_graph(frozen_graph, "some_directory", "my_model.pb", as_text=False)
</code></pre>

blocks|key|4350825|text|freeze_session方法运行良好。但与保存到检查点文件然后使用TensorFlow附带的freeze_graph工具相比，在我看来似乎更简单，因为它更容易维护。您只需执行以下两个步骤：|type|unstyled|depth|inlineStyleRanges|entityRanges|data|4350826|首先，在Keras代码model.fit(...)之后添加，并训练模型：|offset|length|style|CODE|4350827|from+keras+import+backend+as+K
import+tensorflow+as+tf
print(model.output.op.name)
saver+=+tf.train.Saver()
saver.save(K.get_session(),+'/tmp/keras_model.ckpt')|code-block|syntax|javascript|4350828|然后cd到您的TensorFlow根目录，运行：|4350829|python+tensorflow/python/tools/freeze_graph.py+\
--input_meta_graph=/tmp/keras_model.ckpt.meta+\
--input_checkpoint=/tmp/keras_model.ckpt+\
--output_graph=/tmp/keras_frozen.pb+\
--output_node_names="<output_node_name_printed_in_step_1>"+\
--input_binary=true|4350830|entityMap^0|0|B|E|0|0|0|0^^$0|@$1|2|3|4|5|6|7|S|8|@]|9|@]|A|$]]|$1|B|3|C|5|6|7|T|8|@$D|U|E|V|F|G]]|9|@]|A|$]]|$1|H|3|I|5|J|7|W|8|@]|9|@]|A|$K|L]]|$1|M|3|N|5|6|7|X|8|@]|9|@]|A|$]]|$1|O|3|P|5|J|7|Y|8|@]|9|@]|A|$K|L]]|$1|Q|3|-4|5|6|7|Z|8|@]|9|@]|A|$]]]|R|$]]

The freeze_session method works fine. But compared to saving to a checkpoint file then using the freeze_graph tool that comes with TensorFlow seems simpler to me, as it's easier to maintain. All you need to do is the following two steps:

First, add after your Keras code <code>model.fit(...)</code> and train your model:

<pre><code>from keras import backend as K
import tensorflow as tf
print(model.output.op.name)
saver = tf.train.Saver()
saver.save(K.get_session(), '/tmp/keras_model.ckpt')
</code></pre>

Then cd to your TensorFlow root directory, run:

<pre><code>python tensorflow/python/tools/freeze_graph.py \
--input_meta_graph=/tmp/keras_model.ckpt.meta \
--input_checkpoint=/tmp/keras_model.ckpt \
--output_graph=/tmp/keras_frozen.pb \
--output_node_names="&lt;output_node_name_printed_in_step_1&gt;" \
--input_binary=true
</code></pre>

blocks|key|4350845|text|请使用tf.saved_model.simple_save，一些示例代码：|type|unstyled|depth|inlineStyleRanges|entityRanges|offset|length|data|4350846|with+tf.keras.backend.get_session()+as+sess:
++++tf.saved_model.simple_save(
++++++++sess,
++++++++export_path,
++++++++inputs={'input':+keras_model.input},
++++++++outputs={'output':+keras_model.output})|code-block|syntax|javascript|4350847|===update====|4350848|您可以使用as_a_saved_model，示例代码：|4350849|saved_model_path+=+tf.contrib.saved_model.save_keras_model(model,+"./saved_models")|4350850|entityMap|0|LINK|mutability|MUTABLE|url|https://www.tensorflow.org/api_docs/python/tf/saved_model/simple_save|1|https://www.tensorflow.org/api_docs/python/tf/contrib/saved_model/save_keras_model^0|3|Q|0|0|0|0|5|G|1|0|0^^$0|@$1|2|3|4|5|6|7|Y|8|@]|9|@$A|Z|B|10|1|11]]|C|$]]|$1|D|3|E|5|F|7|12|8|@]|9|@]|C|$G|H]]|$1|I|3|J|5|6|7|13|8|@]|9|@]|C|$]]|$1|K|3|L|5|6|7|14|8|@]|9|@$A|15|B|16|1|17]]|C|$]]|$1|M|3|N|5|F|7|18|8|@]|9|@]|C|$G|H]]|$1|O|3|-4|5|6|7|19|8|@]|9|@]|C|$]]]|P|$Q|$5|R|S|T|C|$U|V]]|W|$5|R|S|T|C|$U|X]]]]

Please use <a href="https://www.tensorflow.org/api_docs/python/tf/saved_model/simple_save" rel="nofollow noreferrer">tf.saved_model.simple_save</a>, some example codes:

<pre><code>with tf.keras.backend.get_session() as sess:
 tf.saved_model.simple_save(
 sess,
 export_path,
 inputs={'input': keras_model.input},
 outputs={'output': keras_model.output})
</code></pre>

===update====

You can use <a href="https://www.tensorflow.org/api_docs/python/tf/contrib/saved_model/save_keras_model" rel="nofollow noreferrer">as_a_saved_model</a>, example codes:

<pre><code>saved_model_path = tf.contrib.saved_model.save_keras_model(model, "./saved_models")
</code></pre>

blocks|key|4351359|text|使用estimator.export_savedmodel，我们可以很容易地将h5模型转换为保存的模型。在此处检查文档https://www.tensorflow.org/api_docs/python/tf/estimator/Estimator|type|unstyled|depth|inlineStyleRanges|entityRanges|offset|length|data|4351360|def+prepare_image(image_str_tensor):
++++image_contents+=+tf.read_file(image_str_tensor)
++++image+=+tf.image.decode_jpeg(image_contents,+channels=3)
++++image+=+tf.image.resize_images(image,+[224,+224])
++++image+=+tf.cast(image,+tf.float32)
++++return+preprocess_input(image)

def+serving_input_receiver_fn():
++++input_ph+=+tf.placeholder(tf.string,+shape=[None])
++++images_tensor+=+tf.map_fn(
++++++++++prepare_image,+input_ph,+back_prop=False,+dtype=tf.float32)
++++images_tensor+=+tf.image.convert_image_dtype(images_tensor,+
++++++++++++++++++++++dtype=tf.float32)

++++return+tf.estimator.export.ServingInputReceiver({"input":+images_tensor},+
+++++++++++++{'image_url':+input_ph})

estimator+=+tf.keras.estimator.model_to_estimator(
++++keras_model_path=h5_model_path
)

estimator.export_savedmodel(saved_model_path,+serving_input_receiver_fn=serving_input_receiver_fn)|code-block|syntax|javascript|4351361|entityMap|0|LINK|mutability|MUTABLE|url|https://www.tensorflow.org/api_docs/python/tf/estimator/Estimator^0|1N|1T|0|0|0^^$0|@$1|2|3|4|5|6|7|Q|8|@]|9|@$A|R|B|S|1|T]]|C|$]]|$1|D|3|E|5|F|7|U|8|@]|9|@]|C|$G|H]]|$1|I|3|-4|5|6|7|V|8|@]|9|@]|C|$]]]|J|$K|$5|L|M|N|C|$O|P]]]]

using estimator.export_savedmodel we can easily convert h5 model to saved model.
check doc here <a href="https://www.tensorflow.org/api_docs/python/tf/estimator/Estimator" rel="nofollow noreferrer">https://www.tensorflow.org/api_docs/python/tf/estimator/Estimator</a>

<pre><code>def prepare_image(image_str_tensor):
 image_contents = tf.read_file(image_str_tensor)
 image = tf.image.decode_jpeg(image_contents, channels=3)
 image = tf.image.resize_images(image, [224, 224])
 image = tf.cast(image, tf.float32)
 return preprocess_input(image)

def serving_input_receiver_fn():
 input_ph = tf.placeholder(tf.string, shape=[None])
 images_tensor = tf.map_fn(
 prepare_image, input_ph, back_prop=False, dtype=tf.float32)
 images_tensor = tf.image.convert_image_dtype(images_tensor, 
 dtype=tf.float32)

 return tf.estimator.export.ServingInputReceiver({"input": images_tensor}, 
 {'image_url': input_ph})

estimator = tf.keras.estimator.model_to_estimator(
 keras_model_path=h5_model_path
)

estimator.export_savedmodel(saved_model_path, serving_input_receiver_fn=serving_input_receiver_fn)
</code></pre>

blocks|key|4350897|text|这个解决方案对我很有效。对https://medium.com/tensorflow/training-and-serving-ml-models-with-tf-keras-fd975cc0fa27的礼貌|type|unstyled|depth|inlineStyleRanges|entityRanges|offset|length|data|4350898|import+tensorflow+as+tf

#+The+export+path+contains+the+name+and+the+version+of+the+model
tf.keras.backend.set_learning_phase(0)+#+Ignore+dropout+at+inference
model+=+tf.keras.models.load_model('./model.h5')
export_path+=+'./PlanetModel/1'

#+Fetch+the+Keras+session+and+save+the+model
#+The+signature+definition+is+defined+by+the+input+and+output+tensors
#+And+stored+with+the+default+serving+key
with+tf.keras.backend.get_session()+as+sess:
++++tf.saved_model.simple_save(
++++++++sess,
++++++++export_path,
++++++++inputs={'input_image':+model.input},
++++++++outputs={t.name:t+for+t+in+model.outputs})|code-block|syntax|javascript|4350899|entityMap|0|LINK|mutability|MUTABLE|url|https://medium.com/tensorflow/training-and-serving-ml-models-with-tf-keras-fd975cc0fa27^0|D|2F|0|0|0^^$0|@$1|2|3|4|5|6|7|Q|8|@]|9|@$A|R|B|S|1|T]]|C|$]]|$1|D|3|E|5|F|7|U|8|@]|9|@]|C|$G|H]]|$1|I|3|-4|5|6|7|V|8|@]|9|@]|C|$]]]|J|$K|$5|L|M|N|C|$O|P]]]]

This solution worked for me. 
Courtesy to <a href="https://medium.com/tensorflow/training-and-serving-ml-models-with-tf-keras-fd975cc0fa27" rel="noreferrer">https://medium.com/tensorflow/training-and-serving-ml-models-with-tf-keras-fd975cc0fa27</a>

<pre><code>import tensorflow as tf

# The export path contains the name and the version of the model
tf.keras.backend.set_learning_phase(0) # Ignore dropout at inference
model = tf.keras.models.load_model('./model.h5')
export_path = './PlanetModel/1'

# Fetch the Keras session and save the model
# The signature definition is defined by the input and output tensors
# And stored with the default serving key
with tf.keras.backend.get_session() as sess:
 tf.saved_model.simple_save(
 sess,
 export_path,
 inputs={'input_image': model.input},
 outputs={t.name:t for t in model.outputs})
</code></pre>

blocks|key|4350957|text|tf+2.2.0|type|unstyled|depth|inlineStyleRanges|entityRanges|data|4350958|导入tensorflow.keras而不仅仅是keras，因为它会将模型作为keras.engine.sequential.Sequential对象加载，而不能直接转换为tensorflow+.pb格式|4350959|#import+keras
import+tensorflow.keras+as+keras
model+=+keras.models.load_model(load_path)
model.save(save_path)|code-block|syntax|javascript|4350960|entityMap^0|0|0|0^^$0|@$1|2|3|4|5|6|7|K|8|@]|9|@]|A|$]]|$1|B|3|C|5|6|7|L|8|@]|9|@]|A|$]]|$1|D|3|E|5|F|7|M|8|@]|9|@]|A|$G|H]]|$1|I|3|-4|5|6|7|N|8|@]|9|@]|A|$]]]|J|$]]

tf 2.2.0
import tensorflow.keras instead of just keras, because it will load your model as keras.engine.sequential.Sequential object which cannot be directly convertible into tensorflow .pb format
<pre><code>#import keras
import tensorflow.keras as keras
model = keras.models.load_model(load_path)
model.save(save_path)
</code></pre>

blocks|key|4351026|text|使用tensorflow+2.x：如果您只想在pbtxt中保存图形定义，则使用以下代码。|type|unstyled|depth|inlineStyleRanges|offset|length|style|CODE|entityRanges|data|4351027|import+tensorflow+as+tf
keras_model+=+...
tf.io.write_graph(
++keras_model.output.graph,
++'model_dir',
++'model.pbtxt',
++as_text=True,
)|code-block|syntax|javascript|4351028|entityMap^0|2|E|N|5|0|0^^$0|@$1|2|3|4|5|6|7|M|8|@$9|N|A|O|B|C]|$9|P|A|Q|B|C]]|D|@]|E|$]]|$1|F|3|G|5|H|7|R|8|@]|D|@]|E|$I|J]]|$1|K|3|-4|5|6|7|S|8|@]|D|@]|E|$]]]|L|$]]

With <code>tensorflow 2.x</code>:
If you want to save only the graph definition in <code>pbtxt</code> then use the below code.
<pre class="lang-py prettyprint-override"><code>import tensorflow as tf
keras_model = ...
tf.io.write_graph(
 keras_model.output.graph,
 'model_dir',
 'model.pbtxt',
 as_text=True,
)
</code></pre>

blocks|key|4351535|text|在用户试图将Mask-RCNN模型/权重转换为冻结图形的情况下，这里的大多数答案都不够。|type|unstyled|depth|inlineStyleRanges|entityRanges|data|4351536|可以在mrcnn/model.py文件中保存模型(.h5)权重时执行此操作。只需进行以下更改(git+diff)|offset|length|style|CODE|4351537|%2B++++def+freeze_session(self,+session,+keep_var_names=None,+output_names=None,+clear_devices=True):
%2B++++++++"""
%2B++++++++Freezes+the+state+of+a+session+into+a+pruned+computation+graph.
%2B
%2B++++++++Creates+a+new+computation+graph+where+variable+nodes+are+replaced+by
%2B++++++++constants+taking+their+current+value+in+the+session.+The+new+graph+will+be
%2B++++++++pruned+so+subgraphs+that+are+not+necessary+to+compute+the+requested
%2B++++++++outputs+are+removed.
%2B++++++++@param+session+The+TensorFlow+session+to+be+frozen.
%2B++++++++@param+keep_var_names+A+list+of+variable+names+that+should+not+be+frozen,
%2B++++++++++++++++++++++++++++++or+None+to+freeze+all+the+variables+in+the+graph.
%2B++++++++@param+output_names+Names+of+the+relevant+graph+outputs.
%2B++++++++@param+clear_devices+Remove+the+device+directives+from+the+graph+for+better+portability.
%2B++++++++@return+The+frozen+graph+definition.
%2B++++++++"""
%2B++++++++graph+=+session.graph
%2B++++++++with+graph.as_default():
%2B++++++++++++freeze_var_names+=+list(set(v.op.name+for+v+in+tf.global_variables()).difference(keep_var_names+or+[]))
%2B++++++++++++output_names+=+output_names+or+[]
%2B++++++++++++output_names+%2B=+[v.op.name+for+v+in+tf.global_variables()]
%2B++++++++++++input_graph_def+=+graph.as_graph_def()
%2B++++++++++++if+clear_devices:
%2B++++++++++++++++for+node+in+input_graph_def.node:
%2B++++++++++++++++++++node.device+=+""
%2B++++++++++++frozen_graph+=+tf.graph_util.convert_variables_to_constants(
%2B++++++++++++++++session,+input_graph_def,+output_names,+freeze_var_names)
%2B++++++++++++return+frozen_graph
%2B
+++++def+train(self,+train_dataset,+val_dataset,+learning_rate,+epochs,+layers,
+++++++++++++++augmentation=None,+custom_callbacks=None,+no_augmentation_sources=None):
+++++++++"""Train+the+model.
@@+-2373,6+%2B2401,12+@@+class+MaskRCNN():
+++++++++++++workers=workers,
+++++++++++++use_multiprocessing=True,
+++++++++)
%2B++++++++#######using+session+and+saving+.pb+file##
%2B++++++++frozen_graph+=+self.freeze_session(K.get_session(),
%2B++++++++++++++++++++++++++++++output_names=[out.op.name+for+out+in+self.keras_model.outputs])
%2B++++++++print('\n\n\t\t*******+Writing+Frozen+Graph+in+logs+directory+*******\n\n')
%2B++++++++tf.train.write_graph(frozen_graph,+self.log_dir,+"my_model.pb",+as_text=False)
%2B
+++++++++self.epoch+=+max(self.epoch,+epochs)|code-block|syntax|javascript|4351538|完整的文件可以在HERE上找到。有了它，我能够为coco和imagenet权重转换ResNet50和ResNet101主干。|4351539|entityMap|0|LINK|mutability|MUTABLE|url|https://www.dropbox.com/s/hk282nro37r7pfr/model.py^0|0|3|E|P|3|1B|8|0|0|8|4|0|0^^$0|@$1|2|3|4|5|6|7|W|8|@]|9|@]|A|$]]|$1|B|3|C|5|6|7|X|8|@$D|Y|E|Z|F|G]|$D|10|E|11|F|G]|$D|12|E|13|F|G]]|9|@]|A|$]]|$1|H|3|I|5|J|7|14|8|@]|9|@]|A|$K|L]]|$1|M|3|N|5|6|7|15|8|@]|9|@$D|16|E|17|1|18]]|A|$]]|$1|O|3|-4|5|6|7|19|8|@]|9|@]|A|$]]]|P|$Q|$5|R|S|T|A|$U|V]]]]

In the case of users trying to convert a Mask-RCNN model/weights into a frozen graph, most answers here won't suffice.
This can be done while saving the model (<code>.h5</code>) weights in the <code>mrcnn/model.py</code> file. Just need to make the following changes (<code>git diff</code>)
<pre><code>+ def freeze_session(self, session, keep_var_names=None, output_names=None, clear_devices=True):
+ &quot;&quot;&quot;
+ Freezes the state of a session into a pruned computation graph.
+
+ Creates a new computation graph where variable nodes are replaced by
+ constants taking their current value in the session. The new graph will be
+ pruned so subgraphs that are not necessary to compute the requested
+ outputs are removed.
+ @param session The TensorFlow session to be frozen.
+ @param keep_var_names A list of variable names that should not be frozen,
+ or None to freeze all the variables in the graph.
+ @param output_names Names of the relevant graph outputs.
+ @param clear_devices Remove the device directives from the graph for better portability.
+ @return The frozen graph definition.
+ &quot;&quot;&quot;
+ graph = session.graph
+ with graph.as_default():
+ freeze_var_names = list(set(v.op.name for v in tf.global_variables()).difference(keep_var_names or []))
+ output_names = output_names or []
+ output_names += [v.op.name for v in tf.global_variables()]
+ input_graph_def = graph.as_graph_def()
+ if clear_devices:
+ for node in input_graph_def.node:
+ node.device = &quot;&quot;
+ frozen_graph = tf.graph_util.convert_variables_to_constants(
+ session, input_graph_def, output_names, freeze_var_names)
+ return frozen_graph
+
 def train(self, train_dataset, val_dataset, learning_rate, epochs, layers,
 augmentation=None, custom_callbacks=None, no_augmentation_sources=None):
 &quot;&quot;&quot;Train the model.
@@ -2373,6 +2401,12 @@ class MaskRCNN():
 workers=workers,
 use_multiprocessing=True,
 )
+ #######using session and saving .pb file##
+ frozen_graph = self.freeze_session(K.get_session(),
+ output_names=[out.op.name for out in self.keras_model.outputs])
+ print('\n\n\t\t******* Writing Frozen Graph in logs directory *******\n\n')
+ tf.train.write_graph(frozen_graph, self.log_dir, &quot;my_model.pb&quot;, as_text=False)
+
 self.epoch = max(self.epoch, epochs)

</code></pre>
The complete file can be found <a href="https://www.dropbox.com/s/hk282nro37r7pfr/model.py" rel="nofollow noreferrer">HERE</a>. With it, I was able to convert ResNet50 and ResNet101 backbones for both coco as well as imagenet weights.

I have fine-tuned inception model with a new dataset and saved it as ".h5" model in Keras. now my goal is to run my model on android Tensorflow which accepts ".pb" extension only. question is that is there any library in Keras or tensorflow to do this conversion? I have seen this post so far : <a href="https://blog.keras.io/keras-as-a-simplified-interface-to-tensorflow-tutorial.html" rel="noreferrer">https://blog.keras.io/keras-as-a-simplified-interface-to-tensorflow-tutorial.html</a> but can't figure out yet.

How to export Keras .h5 to tensorflow .pb?

Python

我使用新的数据集对初始模型进行了微调，并将其保存为Keras中的".h5“模型。现在我的目标是在android Tensorflow上运行我的模型，它只接受".pb“扩展。问题是，在Keras或tensorflow中是否有任何库可以执行此转换？到目前为止，我已经看过这篇文章了：，但还不明白。

问如何将Keras .h5导出到tensorflow .pb？
EN

社区

活动

资源

关于

腾讯云开发者

热门产品

热门推荐

更多推荐

问如何将Keras .h5导出到tensorflow .pb？EN

社区

活动

资源

关于

腾讯云开发者

热门产品

热门推荐

更多推荐

问如何将Keras .h5导出到tensorflow .pb？
EN