import fails: "undefined symbol: _ZN10tensorflow12OpDefBuilder4AttrESs"

[moreymat]

I encountered this bug which is most probably a duplicate of #30 that has been closed.
Is it related to #160 (comment) ?

System information

• Have I written custom code (as opposed to using a stock example script provided in TensorFlow): no
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Linux Ubuntu 18.04 LTS
• Mobile device (e.g. iPhone 8, Pixel 2, Samsung Galaxy) if the issue happens on mobile device: no
• TensorFlow installed from (source or binary): binary
• TensorFlow version (use command below): 2.0.0
• Python version: Anaconda python 3.7.5
• CUDA/cuDNN version: None
• GPU model and memory: None

Describe the current behavior
Error on importing tensorflow-text making it impossible to be imported.

Describe the expected behavior
Library can be effortlessly imported and used.

Code to reproduce the issue
Provide a reproducible test case that is the bare minimum necessary to generate the problem.

I created a new minimal environment using

conda create -n tf-test tensorflow python=3.7
conda activate tf-test
pip install tensorflow-text

then, when trying to import tensorflow_text the following error appears

$ python
Python 3.7.5 (default, Oct 25 2019, 15:51:11) 
[GCC 7.3.0] :: Anaconda, Inc. on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import tensorflow as tf
>>> import tensorflow_text as text
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/mathieu/miniconda3/envs/tf-test/lib/python3.7/site-packages/tensorflow_text/__init__.py", line 21, in <module>
    from tensorflow_text.python import metrics
  File "/home/mathieu/miniconda3/envs/tf-test/lib/python3.7/site-packages/tensorflow_text/python/metrics/__init__.py", line 20, in <module>
    from tensorflow_text.python.metrics.text_similarity_metric_ops import *
  File "/home/mathieu/miniconda3/envs/tf-test/lib/python3.7/site-packages/tensorflow_text/python/metrics/text_similarity_metric_ops.py", line 28, in <module>
    gen_text_similarity_metric_ops = load_library.load_op_library(resource_loader.get_path_to_datafile('_text_similarity_metric_ops.so'))
  File "/home/mathieu/miniconda3/envs/tf-test/lib/python3.7/site-packages/tensorflow_core/python/framework/load_library.py", line 61, in load_op_library
    lib_handle = py_tf.TF_LoadLibrary(library_filename)
tensorflow.python.framework.errors_impl.NotFoundError: /home/mathieu/miniconda3/envs/tf-test/lib/python3.7/site-packages/tensorflow_text/python/metrics/_text_similarity_metric_ops.so: undefined symbol: _ZN10tensorflow12OpDefBuilder4AttrESs

[broken]

This looks like a conda issue. If you don't create the conda environment with tensorflow in the beginning it works fine.

conda create -n tf-test python=3.7
conda activate tf-test
pip install tensorflow
pip install tensorflow-text

Do you know what is different about installing it at creation vs using pip after?

[broken]

This is a month old and doesn't appear like a tf text issue, so I'm closing.

[moreymat]

Sorry I could not reply earlier then lost sight of your answer.
This works :

pip install tensorflow
pip install tensorflow-text

but this does not :

conda install tensorflow
pip install tensorflow-text

Could it be an issue with the conda recipe for tensorflow ?

[elliotvilhelm]

Same issue here using conda installed tensorflow 2.0.0 and tensorflow-text==2.0.0

[shailensobhee]

This is a month old and doesn't appear like a tf text issue, so I'm closing.

This doesn't look like a legit reason to close a ticket.

[soumayan]

same issue-
pip install tensorflow-gpu==2.2.0rc1
pip install tensorflow-text==2.2.0rc1
import tensorflow_text

NotFoundError: /opt/conda/lib/python3.6/site-packages/tensorflow_text/python/metrics/_text_similarity_metric_ops.so: undefined symbol: _ZN10tensorflow8OpKernel11TraceStringEPNS_15OpKernelContextEb

[zzj0402]

Same issue in Kaggle notebook. This issue is needed and not resolved. Poor task management.

[soumayan]

@zzj0402 install tensorflow 2.0 in kaggle notebook,then this problem will be removed

[karllessard]

If that can help, I work in TensorFlow Java and and just tried to load directly the _wordpiece_tokenizer.so library extracted from TF-Text 2.2.0 python wheels and link it with libtensorflow_framework.so.2 that we distribute (which is built now from TF 2.2.0). I got the same error as @soumayan , who had to downgrade to TF2.0.0 if I understand correctly:

Exception in thread "main" java.lang.UnsatisfiedLinkError: /home/karl/_wordpiece_tokenizer.so: undefined symbol: _ZN10tensorflow8OpKernel11TraceStringEPNS_15OpKernelContextEb

So it looks like there might be an incompatibility between TF 2.2.0 and TF-Text 2.2.0

[zzj0402]

@zzj0402 install tensorflow 2.0 in kaggle notebook,then this problem will be removed

I solved it by toggle GPU on in Kaggle environment settings.

[phihung]

Similar issue in Mac OSX with tensorflow and tensorflow_text installed with pip inside conda env.

import tensorflow as tf  # OK 2.2.0
import tensorflow_text as text

NotFoundError: dlopen(/miniconda3/envs/tf/lib/python3.7/site-packages/tensorflow_text/python/ops/_normalize_ops.dylib, 6): Symbol not found: ____chkstk_darwin
  Referenced from: /miniconda3/envs/tf/lib/python3.7/site-packages/tensorflow_text/python/ops/_normalize_ops.dylib (which was built for Mac OS X 10.15)
  Expected in: /usr/lib/libSystem.B.dylib
 in /miniconda3/envs/tf/lib/python3.7/site-packages/tensorflow_text/python/ops/_normalize_ops.dylib 

[mikkotuomela]

I have this same problem. tensorflow gets imported without problems, importing tensorflow-text results in op's error message.

tensorflow is built from sources (1.15.3), tensorflow-text is stock binary (1.15.1, tried also 1.15.0rc0).
Platform is Ubuntu 18.04.3 LTS. Python 3.6.9. Using (Mini)conda environment.

tensorflow is built without GPU support and tensorflow-gpu is not installed.

Could there still be a compatibility issue? Interestingly, people seem to have exact same issue with TF 2 as well, but here I am using version 1.15.

Any help or ideas would be greatly appreciated.

[snehankekre]

I saw the same error on Colab. The solution was that the tensorflow-text version needs to match the tf version. Since Colab was using tensorflow==2.2.0, I was successfully able to install and import tensorflow-text after:
!pip install tensorflow-text==2.2.0

[parthlathiya2697]

Solved:
I was performing tensorflow object detection training and encountered this type of error. It got resolved when I installed tensorflow version 2.1.0, above versions still has that error

[adenilso]

I saw the same error on Colab. The solution was that the tensorflow-text version needs to match the tf version. Since Colab was using tensorflow==2.2.0, I was successfully able to install and import tensorflow-text after:
!pip install tensorflow-text==2.2.0

Thank you! I worked for me.

[nershman]

Changing tensorflow version just gives me a different error... Symbol not found: ____chkstk_darwin

In my case I believe the issue is due to new releases being built on OSX 10.15, info here

[broken]

Nobody monitors closed issues, so I doubt anybody on the team has noticed any of these replies. If anybody on this thread is having a problem, please create a new issue with the error, tf version, tf text version, os, and if you are using an environment like conda or pyenv.

[oksanatkach]

I got this error on Google Colab.

After hours of struggle, I am 99% this is some sort of a VM memory or path issue. It looks like you are using a library of one version (if you run pip list), but it's actually a different version.

To fix this, completely disconnect and delete the runtime. If you don't do this, I think some stuff in the memory persists forever, even if you close Colab. Then, when you switch to GPU, you will get a GPU environment with tensorflow version 2.9 installed and no tensorflow-text installed. Separately run pip install with versions specified:

pip install tensorflow==2.10
pip install tensorflow-text==2.10

This works.

If you try to run tensorflow v 2.10 with tensorflow-text v 2.9, you get this error too. This is why I believe it's some sort of pip installation memory issue?

Makes sense that tensorflow don't want to solve this, as it is likely just the Google Colab VM problem.

[pras135]

this worked for me
!pip install -U tensorflow
!pip install -U tensorflow-text
as mentioned here https://stackoverflow.com/a/72964778/5238639