Tensorflow: tensorboard: view graph from saved_model.pb file [feature request]

[brandondutra]

Plot a graph from just a saved_model.pb file.

Currently tensorboard only works given a training folder containing checkpoints and summary events. Understanding the output graph is important, especially if you don't have access to the training output files.

[jubjamie]

@brandondutra As far as I know, you don't need to create any summaries to load the graph into Tensorboard. If you begin to create a summarywriter and then add the graph to it then you should see the graph appear in Tensorboard. I have some code that does that (as part of another project). I do however agree that being able to import something into Tensorboard would be handy.

I've quickly created a bit of code to load a graph into Tensorboard. See how that goes. Also available as a gist here

import tensorflow as tf
from tensorflow.python.platform import gfile
with tf.Session() as sess:
    model_filename ='PATH_TO_PB.pb'
    with gfile.FastGFile(model_filename, 'rb') as f:
        graph_def = tf.GraphDef()
        graph_def.ParseFromString(f.read())
        g_in = tf.import_graph_def(graph_def)
LOGDIR='YOUR_LOG_LOCATION'
train_writer = tf.summary.FileWriter(LOGDIR)
train_writer.add_graph(sess.graph)

You should then be able to see this in Tensorboard.
Something like that perhaps in the mean time? I might put a PR in to get some conversation going,

[brandondutra]

This is awesome! Thank you for your help. I took your script and ran it, but I ran into an issue:

google.protobuf.message.DecodeError: Error parsing message

Because the error is in parsing the message, I looked into tensorflow protos. I modified your script, and got it working on my saved_model.pb file.

import tensorflow as tf
import sys
from tensorflow.python.platform import gfile

from tensorflow.core.protobuf import saved_model_pb2
from tensorflow.python.util import compat

with tf.Session() as sess:
    model_filename ='saved_model.pb'
    with gfile.FastGFile(model_filename, 'rb') as f:

        data = compat.as_bytes(f.read())
        sm = saved_model_pb2.SavedModel()
        sm.ParseFromString(data)
        #print(sm)
        if 1 != len(sm.meta_graphs):
            print('More than one graph found. Not sure which to write')
            sys.exit(1)

        #graph_def = tf.GraphDef()
        #graph_def.ParseFromString(sm.meta_graphs[0])
        g_in = tf.import_graph_def(sm.meta_graphs[0].graph_def)
LOGDIR='YOUR_LOG_LOCATION'
train_writer = tf.summary.FileWriter(LOGDIR)
train_writer.add_graph(sess.graph)

Again, thank you!

[jubjamie]

@brandondutra I guess that there was something odd with your .pb then. Worked fine on mine! Glad I could help.

[aselle]

Closing since this feature is available through existing TensorBoard functionality.

[jubjamie]

@aselle Can I just confirm what you mean by existing Tensorboard functionality? Do you mean the code that myself and @brandondutra wrote (and associated functions) or is there another way?

[aselle]

Yes that's what I mean. @dandelionmane, do you want to add a more direct way to visualize saved graphs?

[brandondutra]

Yes. Telling users to write a python script to visualize a saved model graph is a very bad user experience [it's not even documented as a thing that can be done], and seems like something tensorboard should be very good at.

[passerbyd]

Actually, tf.summary.FileWriter accepts a GraphDef parameter which marks as deprecated.

[brandondutra]

I think the point/ask of this issue is that tensorboard should show the graph from a saved_model.pb file

[MikulasZelinka]

When trying to load a frozen pb model to visualise its graph, the snippets above didn't work for me in tensorflow 1.4, Python 3.6 (the resulting folder was empty).

Using the Upload - Choose file option in tensorboard's GUI gave Graph visualization failed: The graph is empty.

However, using the import_to_tensorboard function did the trick:

https://github.com/tensorflow/tensorflow/blob/r1.4/tensorflow/python/tools/import_pb_to_tensorboard.py

[gustavz]

hey @MikulasZelinka how to use this function?

can you post an example how to call it with an example model and log dir?
For me somehow it does not work

@jubjamie your code works fine, thanks :+1:

Thank you!

[withai]

@GustavZ You can visualize the graph of your frozen pb model using this command

python import_pb_to_tensorboard.py --model_dir="path/to/your/model/file.pb" --log_dir="path/to/your/log/dir"

[Jon889]

I found I had to change @brandondutra script slightly to add a flush and close at the end.
Else on some models nothing would be written to the LOG_DIR.
Also found that import_pb_to_tensorboard gave the same error as google.protobuf.message.DecodeError: Error parsing message so had to use this in the first place.

import tensorflow as tf
import sys
from tensorflow.python.platform import gfile

from tensorflow.core.protobuf import saved_model_pb2
from tensorflow.python.util import compat

with tf.Session() as sess:
    model_filename =sys.argv[1]
    with gfile.FastGFile(model_filename, 'rb') as f:

        data = compat.as_bytes(f.read())
        sm = saved_model_pb2.SavedModel()
        sm.ParseFromString(data)
        #print(sm)
        if 1 != len(sm.meta_graphs):
            print('More than one graph found. Not sure which to write')
            sys.exit(1)

        #graph_def = tf.GraphDef()
        #graph_def.ParseFromString(sm.meta_graphs[0])
        g_in = tf.import_graph_def(sm.meta_graphs[0].graph_def)
    LOGDIR=sys.argv[2]

train_writer = tf.summary.FileWriter(LOGDIR)
train_writer.add_graph(sess.graph)
train_writer.flush()
train_writer.close()

[jhw6976]

@Jon889, I followed you suggestion and I can overcome error "the google.protobuf.message.DecodeError: Error parsing message"

But when I use tensorboard to view the graph, it is empty graph,

i made below change, and i can view graph,
I tested with .pb created by "mnist_saved_model.py" (https://github.com/tensorflow/serving/blob/master/tensorflow_serving/example/mnist_saved_model.py)

code snippet:

import argparse
import sys
import tensorflow as tf

from tensorflow.python.platform import app
from tensorflow.python.summary import summary

def import_to_tensorboard(model_dir, log_dir):
  """View an imported protobuf model (`.pb` file) as a graph in Tensorboard.

  Args:
    model_dir: The location of the protobuf (`pb`) model to visualize
    log_dir: The location for the Tensorboard log to begin visualization from.

  Usage:
    Call this function with your model location and desired log directory.
    Launch Tensorboard by pointing it to the log directory.
    View your imported `.pb` model as a graph.
  """

  with tf.Session(graph=tf.Graph()) as sess:
    tf.saved_model.loader.load(
        sess, [tf.saved_model.tag_constants.SERVING], model_dir)

    pb_visual_writer = summary.FileWriter(log_dir)
    pb_visual_writer.add_graph(sess.graph)
    print("Model Imported. Visualize by running: "
          "tensorboard --logdir={}".format(log_dir))

[cyberphantom]

If you have two .pb files with the same structure and format, is it possible to add them together???

Let say that you have a trained model then you made a fine-tune for new calsses, then you add the result pb file to the original one so you reduce training time and having one pb file like you made a full training operation

[cxx120]

I don't really think import_pb_to_tensorboard solves the issue here: 1) firstly, model_dir takes a graph.pb, instead of a saved_model.pb; 2) what is it with "log_idr"?

On a separate note: I think TF folks could've done a much better job handling RFEs from users, instead of just keeping on closing them.

[bbbacudo]

how tom open .pb files?

[xyzdcgan]

While adding my frozen.pb file to it, it gives me an error
Traceback (most recent call last):
File "convert.py", line 30, in
g_in = tf.import_graph_def(graph_def)
File "/home/ios/.local/lib/python2.7/site-packages/tensorflow/python/util/deprecation.py", line 432, in new_func
return func(args, *kwargs)
File "/home/ios/.local/lib/python2.7/site-packages/tensorflow/python/framework/importer.py", line 671, in import_graph_def
node, 'Input tensor %r %s' % (input_name, te)))
ValueError: graph_def is invalid at node u'Genc/Conv/BatchNorm/AssignMovingAvg': Input tensor 'Genc/Conv/BatchNorm/moving_mean:0' Cannot convert a tensor of type float32 to an input of type float32_ref.

and if run it through another code which is:

import tensorflow as tf
import sys
from tensorflow.python.platform import gfile

from tensorflow.core.protobuf import saved_model_pb2
from tensorflow.python.util import compat

with tf.Session() as sess:
model_filename ='saved_model.pb'
with gfile.FastGFile(model_filename, 'rb') as f:

data = compat.as_bytes(f.read())
sm = saved_model_pb2.SavedModel()
sm.ParseFromString(data)
#print(sm)
if 1 != len(sm.meta_graphs):
    print('More than one graph found. Not sure which to write')
    sys.exit(1)

#graph_def = tf.GraphDef()
#graph_def.ParseFromString(sm.meta_graphs[0])
g_in = tf.import_graph_def(sm.meta_graphs[0].graph_def)

LOGDIR='YOUR_LOG_LOCATION'
train_writer = tf.summary.FileWriter(LOGDIR)
train_writer.add_graph(sess.graph)

and gave me error:

google.protobuf.message.DecodeError: Error parsing message

any solution for this..??

i converted to .pb file from checkpoint file using tensorflow frozen_python file.

[xyzdcgan]

I also tried to run it through import_pb_to_tensorboard.py file but in that also i got an error of float32 to float32_ref

I think there is something wrong in .pb file, but i converted it through frozen_graph.py of tensorflow . and passed all the correct output_node_name. so i dont figure what is error.

[mrgloom]

First answers are outdated:

Now you can use:
https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/tools/import_pb_to_tensorboard.py

[depthwise]

No you can't:

Traceback (most recent call last):
  File "/home/user/.local/lib/python3.6/site-packages/tensorflow_core/python/tools/import_pb_to_tensorboard.py", line 86, in <module>
    app.run(main=main, argv=[sys.argv[0]] + unparsed)
  File "/home/user/.local/lib/python3.6/site-packages/tensorflow_core/python/platform/app.py", line 40, in run
    _run(main=main, argv=argv, flags_parser=_parse_flags_tolerate_undef)
  File "/home/user/.local/lib/python3.6/site-packages/absl/app.py", line 300, in run
    _run_main(main, args)
  File "/home/user/.local/lib/python3.6/site-packages/absl/app.py", line 251, in _run_main
    sys.exit(main(argv))
  File "/home/user/.local/lib/python3.6/site-packages/tensorflow_core/python/tools/import_pb_to_tensorboard.py", line 68, in main
    import_to_tensorboard(FLAGS.model_dir, FLAGS.log_dir)
  File "/home/user/.local/lib/python3.6/site-packages/tensorflow_core/python/tools/import_pb_to_tensorboard.py", line 58, in import_to_tensorboard
    graph_def.ParseFromString(f.read())
google.protobuf.message.DecodeError: Error parsing message

This is in 2.0. The whole SavedModel debacle is an unmitigated disaster.

[luvwinnie]

any solution to this?

[ltn100]

^ the above PR should fix the problem. For those that are impatient:

wget https://raw.githubusercontent.com/ltn100/tensorflow/feature/LN/fix_import_pb_to_tensorboard/tensorflow/python/tools/import_pb_to_tensorboard.py
python import_pb_to_tensorboard.py --model_dir model_dir/1 --log_dir logs/001

[tsvetiko]

@brandondutra As far as I know, you don't need to create any summaries to load the graph into Tensorboard. If you begin to create a summarywriter and then add the graph to it then you should see the graph appear in Tensorboard. I have some code that does that (as part of another project). I do however agree that being able to import something into Tensorboard would be handy.

I've quickly created a bit of code to load a graph into Tensorboard. See how that goes. Also available as a gist here

import tensorflow as tf
from tensorflow.python.platform import gfile
with tf.Session() as sess:
    model_filename ='PATH_TO_PB.pb'
    with gfile.FastGFile(model_filename, 'rb') as f:
        graph_def = tf.GraphDef()
        graph_def.ParseFromString(f.read())
        g_in = tf.import_graph_def(graph_def)
LOGDIR='YOUR_LOG_LOCATION'
train_writer = tf.summary.FileWriter(LOGDIR)
train_writer.add_graph(sess.graph)

You should then be able to see this in Tensorboard.
Something like that perhaps in the mean time? I might put a PR in to get some conversation going,

Don't forget to flush ;)

train_writer.flush()

[sayakpaul]

Here's a handy Colab notebook in case anyone wants to try out this: https://colab.research.google.com/drive/13LAUUT9tEH2XeoNA_z9A7uE5omc-lzNv

[Arnold1]

@sayakpaul thanks for the notebook!
@jubjamie
@Jon889

any idea why that doesn't work like that - without your function?

from tensorflow.python.tools.import_pb_to_tensorboard import import_to_tensorboard
import os

model_dir = './model'
model = os.path.join(model_dir, 'saved_model.pb')
print(model)
import_to_tensorboard(model_dir=model, log_dir='./logs/')

I see this error:

~$ python main.py 
2020-03-06 23:14:33.962070: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2020-03-06 23:14:33.970833: I tensorflow/core/platform/profile_utils/cpu_utils.cc:94] CPU Frequency: 2200075000 Hz
2020-03-06 23:14:33.972744: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x563e6e9a9f60 initialized for platform Host (this does not guarantee that XLA will be used). Devices:
2020-03-06 23:14:33.972774: I tensorflow/compiler/xla/service/service.cc:176]   StreamExecutor device (0): Host, Default Version
Traceback (most recent call last):
  File "main4.py", line 7, in <module>
    import_to_tensorboard(model_dir=model, log_dir='./logs/')
  File "/home/mike18/venv/lib/python3.6/site-packages/tensorflow_core/python/tools/import_pb_to_tensorboard.py", line 58, in import_to_tensorboard
    graph_def.ParseFromString(f.read())
google.protobuf.message.DecodeError: Error parsing message

also why does this code not work with Tensorflow 2?
https://github.com/tobegit3hub/tfmodel/blob/22c1c48a7049a9cac8e533e317f41697fb85ff7a/tfmodel/savedmodel_analyst.py

when I use your function it works:

# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ================================
"""Imports a protobuf model as a graph in Tensorboard."""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import argparse
import sys

from tensorflow.python.client import session
from tensorflow.python.framework import importer
from tensorflow.python.framework import ops
from tensorflow.python.platform import app
from tensorflow.python.summary import summary
from tensorflow.python.tools import saved_model_utils

# Try importing TensorRT ops if available
# TODO(aaroey): ideally we should import everything from contrib, but currently
# tensorrt module would cause build errors when being imported in
# tensorflow/contrib/__init__.py. Fix it.
# pylint: disable=unused-import,g-import-not-at-top,wildcard-import
try:
  from tensorflow.contrib.tensorrt.ops.gen_trt_engine_op import *
except ImportError:
  pass
# pylint: enable=unused-import,g-import-not-at-top,wildcard-import

def import_to_tensorboard(model_dir, log_dir, tag_set):
  """View an imported protobuf model (`.pb` file) as a graph in Tensorboard.

  Args:
    model_dir: The location of the protobuf (`pb`) model to visualize
    log_dir: The location for the Tensorboard log to begin visualization from.
    tag_set: Group of tag(s) of the MetaGraphDef to load, in string format,
        separated by ','. For tag-set contains multiple tags, all tags must be
        passed in.

  Usage:
    Call this function with your model location and desired log directory.
    Launch Tensorboard by pointing it to the log directory.
    View your imported `.pb` model as a graph.
  """
  with session.Session(graph=ops.Graph()) as sess:
    input_graph_def = saved_model_utils.get_meta_graph_def(
        model_dir, tag_set).graph_def
    importer.import_graph_def(input_graph_def)

    pb_visual_writer = summary.FileWriter(log_dir)
    pb_visual_writer.add_graph(sess.graph)
    print("Model Imported. Visualize by running: "
          "tensorboard --logdir={}".format(log_dir))


def main(_):
  import_to_tensorboard(FLAGS.model_dir, FLAGS.log_dir, FLAGS.tag_set)

if __name__ == "__main__":
  parser = argparse.ArgumentParser()
  parser.register("type", "bool", lambda v: v.lower() == "true")
  parser.add_argument(
      "--model_dir",
      type=str,
      default="",
      required=True,
      help="The directory containing the SavedModel to import.")
  parser.add_argument(
      "--log_dir",
      type=str,
      default="",
      required=True,
      help="The location for the Tensorboard log to begin visualization from.")
  parser.add_argument(
      '--tag_set',
      type=str,
      default="serve",
      required=False,
      help='tag-set of graph in SavedModel to load, separated by \',\'')
  FLAGS, unparsed = parser.parse_known_args()
  app.run(main=main, argv=[sys.argv[0]] + unparsed)

and than run:

$ python import_pb_to_tensorboard.py --model_dir ./model --log_dir logs/
2020-03-06 23:23:07.019262: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2020-03-06 23:23:07.027290: I tensorflow/core/platform/profile_utils/cpu_utils.cc:94] CPU Frequency: 2200075000 Hz
2020-03-06 23:23:07.029262: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x557ede9a02e0 initialized for platform Host (this does not guarantee that XLA will be used). Devices:
2020-03-06 23:23:07.029284: I tensorflow/compiler/xla/service/service.cc:176]   StreamExecutor device (0): Host, Default Version
Model Imported. Visualize by running: tensorboard --logdir=logs/

[clockfly]

@sayakpaul thanks for the notebook!
@jubjamie
@Jon889

any idea why that doesn't work like that - without your function?

from tensorflow.python.tools.import_pb_to_tensorboard import import_to_tensorboard
import os

model_dir = './model'
model = os.path.join(model_dir, 'saved_model.pb')
print(model)
import_to_tensorboard(model_dir=model, log_dir='./logs/')

I see this error:

~$ python main.py 
2020-03-06 23:14:33.962070: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2020-03-06 23:14:33.970833: I tensorflow/core/platform/profile_utils/cpu_utils.cc:94] CPU Frequency: 2200075000 Hz
2020-03-06 23:14:33.972744: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x563e6e9a9f60 initialized for platform Host (this does not guarantee that XLA will be used). Devices:
2020-03-06 23:14:33.972774: I tensorflow/compiler/xla/service/service.cc:176]   StreamExecutor device (0): Host, Default Version
Traceback (most recent call last):
  File "main4.py", line 7, in <module>
    import_to_tensorboard(model_dir=model, log_dir='./logs/')
  File "/home/mike18/venv/lib/python3.6/site-packages/tensorflow_core/python/tools/import_pb_to_tensorboard.py", line 58, in import_to_tensorboard
    graph_def.ParseFromString(f.read())
google.protobuf.message.DecodeError: Error parsing message

also why does this code not work with Tensorflow 2?
https://github.com/tobegit3hub/tfmodel/blob/22c1c48a7049a9cac8e533e317f41697fb85ff7a/tfmodel/savedmodel_analyst.py

when I use your function it works:

# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ================================
"""Imports a protobuf model as a graph in Tensorboard."""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import argparse
import sys

from tensorflow.python.client import session
from tensorflow.python.framework import importer
from tensorflow.python.framework import ops
from tensorflow.python.platform import app
from tensorflow.python.summary import summary
from tensorflow.python.tools import saved_model_utils

# Try importing TensorRT ops if available
# TODO(aaroey): ideally we should import everything from contrib, but currently
# tensorrt module would cause build errors when being imported in
# tensorflow/contrib/__init__.py. Fix it.
# pylint: disable=unused-import,g-import-not-at-top,wildcard-import
try:
  from tensorflow.contrib.tensorrt.ops.gen_trt_engine_op import *
except ImportError:
  pass
# pylint: enable=unused-import,g-import-not-at-top,wildcard-import

def import_to_tensorboard(model_dir, log_dir, tag_set):
  """View an imported protobuf model (`.pb` file) as a graph in Tensorboard.

  Args:
    model_dir: The location of the protobuf (`pb`) model to visualize
    log_dir: The location for the Tensorboard log to begin visualization from.
    tag_set: Group of tag(s) of the MetaGraphDef to load, in string format,
        separated by ','. For tag-set contains multiple tags, all tags must be
        passed in.

  Usage:
    Call this function with your model location and desired log directory.
    Launch Tensorboard by pointing it to the log directory.
    View your imported `.pb` model as a graph.
  """
  with session.Session(graph=ops.Graph()) as sess:
    input_graph_def = saved_model_utils.get_meta_graph_def(
        model_dir, tag_set).graph_def
    importer.import_graph_def(input_graph_def)

    pb_visual_writer = summary.FileWriter(log_dir)
    pb_visual_writer.add_graph(sess.graph)
    print("Model Imported. Visualize by running: "
          "tensorboard --logdir={}".format(log_dir))


def main(_):
  import_to_tensorboard(FLAGS.model_dir, FLAGS.log_dir, FLAGS.tag_set)

if __name__ == "__main__":
  parser = argparse.ArgumentParser()
  parser.register("type", "bool", lambda v: v.lower() == "true")
  parser.add_argument(
      "--model_dir",
      type=str,
      default="",
      required=True,
      help="The directory containing the SavedModel to import.")
  parser.add_argument(
      "--log_dir",
      type=str,
      default="",
      required=True,
      help="The location for the Tensorboard log to begin visualization from.")
  parser.add_argument(
      '--tag_set',
      type=str,
      default="serve",
      required=False,
      help='tag-set of graph in SavedModel to load, separated by \',\'')
  FLAGS, unparsed = parser.parse_known_args()
  app.run(main=main, argv=[sys.argv[0]] + unparsed)

and than run:

$ python import_pb_to_tensorboard.py --model_dir ./model --log_dir logs/
2020-03-06 23:23:07.019262: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2020-03-06 23:23:07.027290: I tensorflow/core/platform/profile_utils/cpu_utils.cc:94] CPU Frequency: 2200075000 Hz
2020-03-06 23:23:07.029262: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x557ede9a02e0 initialized for platform Host (this does not guarantee that XLA will be used). Devices:
2020-03-06 23:23:07.029284: I tensorflow/compiler/xla/service/service.cc:176]   StreamExecutor device (0): Host, Default Version
Model Imported. Visualize by running: tensorboard --logdir=logs/

Works for me.

[llealgt]

Here's a handy Colab notebook in case anyone wants to try out this: https://colab.research.google.com/drive/13LAUUT9tEH2XeoNA_z9A7uE5omc-lzNv

I tried the coloab notebook but the graph in tensorboard is not showing anything, is it working for you?