Arg to turn off in-notebook window

[victorflorintsev]

There should be an optional argument in the "%tensorboard" magic (maybe --no-window or --nowindow) to optionally turn off the in-notebook browser window just in case it's impossible for the browser to access TensorBoard directly to prevent outputs like this in notebooks:

Just a nice-to-have feature. Should also appear in the --helpfull prompt.

[wchargin]

Hi @victorflorintsev! Which version of TensorBoard are you using? As of
TensorBoard 1.15.0 or 2.0.0, you should be able to use %tensorboard to
connect to TensorBoard instances even if your Jupyter notebook is on a
remote host, thanks to #2407. If you’re on TensorBoard 2.0 or higher,
you’ll need to pass --bind_all to serve to the network (which was the
default in 1.x but changed in #2589). If you’re on 1.14 or earlier,
perhaps try upgrading and see if that fixes the problem?

Ideally, the steps above should suffice to enable you to actually view
TensorBoard inline rather than just hiding the error window. If you
really do want to use %tensorboard to launch TensorBoard but not
display it, I’d be interested to hear about it and to learn why it
doesn’t suffice to run !tensorboard ... >/dev/null 2>&1 &. It’s not
straightforward to add a special flag used by %tensorboard only,
because the flags are passed down to the main tensorboard binary (as a
design goal, for transparency), but we could perhaps implement this as
an optional argument to notebook.start:

import tensorboard
# something like:
tensorboard.notebook.start("--logdir path/to/tensorboard", skip_display=True)

[victorflorintsev]

@wchargin I am using 2.0.0 but have limited access on the instance we use so I get an error on --bind_all. The only reason I'd prefer the magic is to keep the code clean and self-explanatory. The notebook.start solution is fine I think.

[wchargin]

have limited access on the instance we use so I get an error on
--bind_all

Okay, that’s fair enough. We probably won’t have the bandwidth to get to
this in the near future, but I’d be happy to review a PR that adds that
argument to notebook.start.

[victorflorintsev]

@wchargin May I take a crack at this one?

[wchargin]

Sure, please feel free! The code is in tensorboard/notebook.py here:

tensorboard/tensorboard/notebook.py

Line 123 in 0981d8a

def start(args_string):

Because of the nature of this module, it doesn’t have any automated
tests. To test your changes manually, you can do the following:

1. Run bazel run //tensorboard/pip_package:extract_pip_package, which
   will build the TensorBoard Pip package and copy it to a directory
   whose name is printed out to the console.
2. Create a new virtualenv, and install Jupyter and TensorFlow.
3. In your virtualenv, uninstall any versions of tensorboard or
   tb-nightly, and then directly pip install the local wheel built in
   step 1 above.
4. Run jupyter notebook and perform any manual tests.

As you continue iterating, you don’t strictly need to create a new
virtualenv each time as long as you always make sure to uninstall
TensorBoard before installing the newly built wheel, and building the
Pip package itself is fast enough (about 4 seconds on my machine), so
this doesn’t end up being too much trouble.