"undefined symbol" when compiling the example Op

[idofr]

• Have I written custom code (as opposed to using a stock example script provided in TensorFlow)?: No
• TensorFlow installed from (source or binary)?: binary
• TensorFlow version: 1.0.1
• Bazel version (if compiling from source): NA
• CUDA/cuDNN version: NA
• GPU Model and Memory: NA
• Exact command to reproduce:

Describe the problem clearly

Trying to build and compile the example in "adding a new op"
https://www.tensorflow.org/extend/adding_an_op
This was moved to
tensorflow/examples/adding_an_op
and was not updated in the tutorial page.
After compiling the zero_out_op_kernel_1.cc example, I'm trying to run the zero_out_1_test.py. Yet this returns the following error

File "zero_out_1_test.py", line 25, in
import zero_out_op_1
File "/home/me/tf_compile/tensorflow/tensorflow/examples/adding_an_op/zero_out_op_1.py", line 26, in
'zero_out_op_kernel_1.so'))
File "/usr/local/lib/python3.5/dist-packages/tensorflow/python/framework/load_library.py", line 64, in load_op_library
None, None, error_msg, error_code)
tensorflow.python.framework.errors_impl.NotFoundError: /home/me/tf_compile/tensorflow/tensorflow/examples/adding_an_op/zero_out_op_kernel_1.so: undefined symbol: _ZN10tensorflow8internal21CheckOpMessageBuilder9NewStringB5cxx11Ev

Any ideas what could that symbol be?

[jart]

I tried the example with 1.0.1 and it worked for me. You might be using an unsupported operating system which has a weird version of GCC. We support Ubuntu Linux 14 LTE to 16. Mac OS X El Capitan. CentOS 7 and above.

[idofr]

I'm using Ubuntu 16 with GCC 5.
Could it be that one of the headers was not compiled correctly?

[jart]

Interesting. So you installed the TensorFlow binary distro, which was probably built using GCC 4.x. Then compiled a custom op using GCC 5.x. Out of curiosity, could you try apt-get installing the GCC 4.x series and see if it works? If so, then I might want to bring to the attention of some other people that a binary compatibility problem exists.

[idofr]

Sure, thanks for the quick response. I'll give it a try either tonight or tomorrow at some point and will let you know

[shashankneocfc]

I've the exact same problem : ' undefined symbol: _ZN10tensorflow8internal21CheckOpMessageBuilder9NewStringEv' .

Does '_ZN10tensorflow8internal21CheckOpMessageBuilder9NewStringEv' have a different error meaning with respect to '_ZN10tensorflow8internal21CheckOpMessageBuilder9NewStringB5cxx11Ev' ?

We've defined our own custom op and our code works successfully with tensorflow 0.12.1 and it crashes with the above error with tensorflow 1.0.0. I'm using Ubuntu 16 with gcc version 5.4.
@idofr Did your problem get solved with gcc 4 ?

Edit:
Following link solved my problem :
deepsense-ai/roi-pooling#1

[jart]

@shashankneocfc Thanks for reporting back with the solution.

It turns out we have the magic incantation -D_GLIBCXX_USE_CXX11_ABI=0 in our pip installation and adding an op documentation. Those error messages are really worrisome, but I'm not sure what we can do to make this easier for users.

[Afzal786]

I am facing this problem while adding a custom ops.

lib_handle = py_tf.TF_LoadLibrary(library_filename)
tensorflow.python.framework.errors_impl.NotFoundError: /home/afzal786/PycharmProjects/multi/venv/lib/python3.6/site-packages/multidim_image_augmentation/python/ops/_augmentation_ops.so: undefined symbol: _ZN10tensorflow8internal21CheckOpMessageBuilder9NewStringB5cxx11Ev

undefined symbol: _ZN10tensorflow8internal21CheckOpMessageBuilder9NewStringB5cxx11Ev
gcc and ubuntu version:
gcc (Ubuntu 7.5.0-3ubuntu1~18.04) 7.5.0

TensorFlow version 2.2.0