Do not load WIT TB plugin in TB versions earlier than 2.2

[jameswex]

See tensorflow/tensorboard#3460 for root cause.

This PR ensures that the tensorboard-plugin-wit package will only load the plugin in TensorBoards 2.2 or later, when WIT was removed as a built-in plugin.

[wchargin]

LGTM: a couple comments inline, but all optional.

[wchargin]

Optional, but consider using pkg_resources.parse_version here:

import pkg_resources

_MIN_TENSORBOARD_VERSION = pkg_resources.parse_version("2.2.0")

version = pkg_resources.parse_version(tensorboard.__version__)
if version < _MIN_TENSORBOARD_VERSION:
    return None

This way, you don’t have to do the parsing yourself, and you don’t have
to worry about edge cases (like, 2.9a0 is a valid version number, even
if we’re not super likely to use it).

[jameswex]

ah great didn't know about that build-in lib. tried out packaging but that required a new dep and i didn't want to add one.

used this code and retested.

[wchargin]

Note: tensorboard.__version__ was added in TensorBoard 1.14.0. If
you want to be back-compatible with older versions, you can use:

import tensorboard.version

tb_version = tensorboard.version.VERSION

[jameswex]

don't think we'll need backwards compat with 1.14, as colab falls back to 1.15 with the tensorflow 1.x magic and anyone installing 1.14 from scratch won't get this pip package

[wchargin]

I’ve tested this in Colab with:

%tensorflow_version 1.x

!wget -q 'https://raw.githubusercontent.com/PAIR-code/what-if-tool/a0b339e66603c3e5eeac60b7f064849cfe446923/tensorboard_plugin_wit/wit_plugin_loader.py' -O patch.py

import tensorboard_plugin_wit.wit_plugin_loader
import shutil
shutil.copyfile("patch.py", tensorboard_plugin_wit.wit_plugin_loader.__file__)

%load_ext tensorboard
%tensorboard --logdir logs

and verified that the result launches with a working What-If Tool both
with and without the %tensorflow_version 1.x directive.

[jameswex]

thanks @wchargin , made your pkg_resources change and retested to verify.

[tolga-b]

Looks good to me as well.
Do we need to add pkg_resources as a dependency to pip? A quick search suggests it's part of python setuptools.

[jameswex]

Good quesion. Since this is a TB package that runs as part of TB, and TB has a dep on setuptools, is that situation appropriate for this package? Or should we explicitly also have a requirement on setuptools? @wchargin your opinion?

[wchargin]

Good quesion. Since this is a TB package that runs as part of TB, and TB has a dep on setuptools, is that situation appropriate for this package? Or should we explicitly also have a requirement on setuptools? @wchargin your opinion?

Hi, sorry, just saw this. It might be a little nicer to have an explicit
requirement on setuptools, but I don’t think it’s a big deal either
way, because setuptools is one of the few packages installed by
default in every virtualenv. I actually don’t know how to create a
Python environment that does not have it installed.