Tf-pose-estimation: KeyError: "The name 'TfPoseEstimator/image:0' refers to a Tensor which does not exist. The operation, 'TfPoseEstimator/image', does not exist in the graph."

[OldSaltyFishGo]

how to solve it?thanks.

[nvnnghia]

hi @OldSaltyFishGo I got the same error. Did you know how to fix it? Thanks

Update: After I retrained the cmu network, input node (image) were renamed to split. Therefore, 'TfPoseEstimator/image', does not exist in the graph. The solution is to change TfPoseEstimator/image in estimator.py to TfPoseEstimator/split.

[quchang3]

I got the same error. Did you know how to fix it? Thanks @OldSaltyFishGo @nvnnghia

[nvnnghia]

hi @quchang3 The solution is to change TfPoseEstimator/image in estimator.py to TfPoseEstimator/split

[quchang3]

Thanks for your reply.
When I retrained the CMU,I get checkpoints like this ,meta,ckpt,data. Before I run eval.py, I want to get graph.pb.
then i run run_checkpoint.py, and the network.py is

if sess_for_load is not None:
    if type == 'cmu' or type == 'vgg':
        if not os.path.isfile(pretrain_path_full):
            raise Exception('Model file doesn\'t exist, path=%s' % pretrain_path_full)
       net.load(os.path.join(_get_base_path(), pretrain_path), sess_for_load)
    else:
        s = '%dx%d' % (placeholder_input.shape[2], placeholder_input.shape[1])
        ckpts = {
            'mobilenet': 'trained/mobilenet_%s/model-246038' % s,
            'mobilenet_thin': 'trained/mobilenet_thin_%s/model-449003' % s,
            'mobilenet_fast': 'trained/mobilenet_fast_%s/model-189000' % s,
            'mobilenet_accurate': 'trained/mobilenet_accurate/model-170000'
        }

how to connect the checkpoint with pb , a newly numpy/openpose_coco.npy??

I have trouble in generating the pb with CMU model . could you give me a example??thanks. @nvnnghia

[nvnnghia]

Put this code in your model dir and run.

`import tensorflow as tf

meta_path = 'model-18303.meta' # Your .meta file
output_node_names = ['Openpose/concat_stage7'] # Output nodes

with tf.Session() as sess:

# Restore the graph
saver = tf.train.import_meta_graph(meta_path)

# Load weights
saver.restore(sess,tf.train.latest_checkpoint('.'))

# Freeze the graph
frozen_graph_def = tf.graph_util.convert_variables_to_constants(
    sess,
    sess.graph_def,
    output_node_names)

# Save the frozen graph
with open('output_graph.pb', 'wb') as f:
  f.write(frozen_graph_def.SerializeToString())`

[CompliceU]

Hello, I downloaded the model of cmu from the Internet, I don’t know why this error occurs.@nvnnghia @quchang3

[hundanLi]

I am also meeting this issue. Not solved yet. It seems that the cmu graph doesn't work. Then I just change model to mobilenet_thin or the others. Just like this. It works.

$ roslaunch tfpose_ros demo_video.launch model:=mobilenet_thin

[alexsokolowska]

I've had the same issue on Ubuntu 18.04, even though the code was running perfectly fine on OSX with the exact same versions of dependencies.
I pinned down that the issue was manifesting itself in estimator.py's

` tf.import_graph_def(graph_def,name='TfPoseEstimator')
self.persistent_sess = tf.compat.v1.Session(graph=self.graph, config=tf_config)

    for ts in [n.name for n in tf.compat.v1.get_default_graph().as_graph_def().node]:
        print(ts)

    self.tensor_image = self.graph.get_tensor_by_name('TfPoseEstimator/image:0')
    self.tensor_output = self.graph.get_tensor_by_name('TfPoseEstimator/Openpose/concat_stage7:0')`

The output of print(ts) showed that the name 'TfPoseEstimator' was not appended to the graph correctly. As a result of that, the lines "get_tensor_by_name" could not have matched with any existing tensor.

Deleting the name also failed. In that case, another error occurs:
"TypeError: An op outside of the function building code is being passed a "Graph" tensor. It is possible to have Graph tensors leak out of the function building context by including atf.init_scope in your function building code."

A suggestion from tensorflow github issue worked for me - just paste the following under where you import tensorflow to disable eager execution.

tf.compat.v1.disable_eager_execution()

[longqianh]

Hello! I met the same problem but solved this by simply changing graph.get_tensor_by_name('TfPoseEstimator/image:0') to graph.get_tensor_by_name(image:0'). Hope this could help :)

[mohit9949]

Hello! I met the same problem but solved this by simply changing graph.get_tensor_by_name('TfPoseEstimator/image:0') to graph.get_tensor_by_name(image:0'). Hope this could help :)

F:\Position Estimation\tf-pose-estimation\tf_pose\estimator.py in __init__(self, graph_path, target_size, tf_config, trt_bool)
    338             print(ts)
    339         self.tensor_image = self.graph.get_tensor_by_name('image :0')
--> 340         self.tensor_output = self.graph.get_tensor_by_name('Openpose/concat_stage7:0')
    341         self.tensor_heatMat = self.tensor_output[:, :, :, :19]
    342         self.tensor_pafMat = self.tensor_output[:, :, :, 19:] 

Hello, I tried as you said but its still giving me the same error.
"The name 'TfPoseEstimator/image:0' refers to a Tensor which does not exist. The operation, 'TfPoseEstimator/image', does not exist in the graph."

[Papageno2]

I've had the same issue on Ubuntu 18.04, even though the code was running perfectly fine on OSX with the exact same versions of dependencies.
I pinned down that the issue was manifesting itself in estimator.py's

` tf.import_graph_def(graph_def,name='TfPoseEstimator')
self.persistent_sess = tf.compat.v1.Session(graph=self.graph, config=tf_config)

    for ts in [n.name for n in tf.compat.v1.get_default_graph().as_graph_def().node]:
        print(ts)

    self.tensor_image = self.graph.get_tensor_by_name('TfPoseEstimator/image:0')
    self.tensor_output = self.graph.get_tensor_by_name('TfPoseEstimator/Openpose/concat_stage7:0')`

The output of print(ts) showed that the name 'TfPoseEstimator' was not appended to the graph correctly. As a result of that, the lines "get_tensor_by_name" could not have matched with any existing tensor.

Deleting the name also failed. In that case, another error occurs:
"TypeError: An op outside of the function building code is being passed a "Graph" tensor. It is possible to have Graph tensors leak out of the function building context by including atf.init_scope in your function building code."

A suggestion from tensorflow github issue worked for me - just paste the following under where you import tensorflow to disable eager execution.

tf.compat.v1.disable_eager_execution()

This works for me !
Thx for saving the time!

I try to modify L337 :

        self.tensor_image = self.graph.get_tensor_by_name('TfPoseEstimator/image:0')
       self.tensor_output = self.graph.get_tensor_by_name('TfPoseEstimator/Openpose/concat_stage7:0')

to

        self.tensor_image = self.graph.get_tensor_by_name('image:0')
        self.tensor_output = self.graph.get_tensor_by_name('Openpose/concat_stage7:0')

but new problem occurs....

Only by adding at the begging of estimator.py
tf.compat.v1.disable_eager_execution()

everythings works well !!!!

[yashasvimisra2798]

I am having this issue when I try to run it on google colab. Please help!

[azmathmoosa]

I used https://netron.app/ to visualize the frozen_graph.pb and there was no image node in the graph. As @nvnnghia suggested, I then renamed image:0 to split:0 and it worked.