Tensorflow 2.18 support

[clementpoiret]

Hi all,

Tensorflow 2.17.1 is currently supported by tfio, but tf 2.18 is throwing the following error:

>>> import tensorflow as tf
>>> tf.__version___
'2.18.0
>>> import tensorflow_io as tfio
.devenv/state/venv/lib/python3.11/site-packages/tensorflow_io/python/ops/__init__.py:98: UserWarning: unable to load libtensorflow_io_plugins.so: unable to open file: libtensorflow_io_plugins.so, from paths: ['.devenv/state/venv/lib/python3.11/site-packages/tensorflow_io/python/ops/libtensorflow_io_plugins.so']
caused by: ['.devenv/state/venv/lib/python3.11/site-packages/tensorflow_io/python/ops/libtensorflow_io_plugins.so: undefined symbol: _ZN3tsl8str_util8EndsWithESt17basic_string_viewIcSt11char_traitsIcEES4_']
  warnings.warn(f"unable to load libtensorflow_io_plugins.so: {e}")
.devenv/state/venv/lib/python3.11/site-packages/tensorflow_io/python/ops/__init__.py:104: UserWarning: file system plugins are not loaded: unable to open file: libtensorflow_io.so, from paths: ['.devenv/state/venv/lib/python3.11/site-packages/tensorflow_io/python/ops/libtensorflow_io.so']
caused by: ['.devenv/state/venv/lib/python3.11/site-packages/tensorflow_io/python/ops/libtensorflow_io.so: undefined symbol: _ZN3tsl8str_util9LowercaseB5cxx11ESt17basic_string_viewIcSt11char_traitsIcEE']
  warnings.warn(f"file system plugins are not loaded: {e}")

Pinning tf to 2.17.1 solves the issue.
Is tf 2.18.0 support planned?

[PhysicallyActive]

I second this! Particularly, because this would also make it possible to finally use numpy ^2.0.

[cdeepali]

Even I am looking forward to TF v2.18 support.

[grofte]

I know Tensorflow is on version 2.19 but if you go to the website they show you 2.16.1 API documentation by default. If you choose "versions" you can get the docs for 2.17 but not more recent than that. So I think it's fair that tensorflow-io doesn't support those two versions yet.

[mihaimaruseac]

That's a bug in the TF documentation website. The release process used to have a step to update the documentation but that is currently broken. @learning-to-play 's team knows about it (notified via a bug in tensorflow/tensorflow)

[learning-to-play]

I informed the team responsible for updating the docs multiple weeks ago. I have also followed up with them.

@random-forests @monicadsong

[cdeepali]

Does TF-io support TF v2.18 now?

[mihaimaruseac]

There has been no new release of TF-io since this issue was started

[xin3liang]

Met the same issue when running the ML Perf Storage v2.0 benchmark.
It looks like the current tensorflow-io 0.37.1 doesn't work with tensorflow 2.18+(including 2.18).
The workaround is pip install tensorflow==2.17.1

[iamthebot]

TF 2.17.x doesn't support numpy 2.x though.

[mihaimaruseac]

Latest release of TF made TF-io optional, it's no longer needed.

[iamthebot]

@mihaimaruseac so the s3 support from tf-io is baked into TF?

[mihaimaruseac]

No.

[iamthebot]

Then it's not true at all that "it's not needed"

[mihaimaruseac]

It's no longer needed for pip install tensorflow.

If you want exotic filesystem support, given this repo is stale, you can fork it, add the support, or just use curl. Possibilities exist, this is OSS.

It might even be possible to hack around and install latest tensorflow and tensorflow-io even though they are incompatible, but for the task at hand they could work. Or split the task into two stages running in separate environments.