Support ignoring certain directories under the logdir

[wchargin]

Migrated from tensorflow/tensorflow#7385.

[cipri-tom]

so, any plans to support this ? :) it would be very welcome

I had a look at the code and I think it doesn't require too much knowledge of the inner workings. I think this part, where it takes a path_filter, is the relevant one. My confusion is that it is called from both event_accumulator.py and the plugin_event_accumulator.

So, I would be happy to contribute this small feature, provided that someone from the team can:

1. decide on which of the previous approaches to take
2. explain the difference between the event_accumulator and its corresponding plugin_...

To re-state the options:

1. If there is special character(like #) in folder name, just ignore it.
2. Like git, .ignore file manage folders to ignore.
3. If .nolog file in the folder to ignore.(Like .nomedia file)

[eliorc]

This feature would help a lot.

Once you get to over 15 different models, the tensorboard takes too long to load and fails to load all the logs often.
Also the colors cycle start to get confusing

[stefan-falk]

Totally agree. :) Wouldn't hurt if one could just use wildcards e.g.

tensorboard --logdir modelname_*/ 

[gabegrand]

Any plans to support this feature in 2019? It would be most welcome.

[fuzihaofzh]

Also need this feature.

[shtratos]

Here's a workaround for specifying path patterns with some bash-fu and a "discouraged" --logdir_spec option which supports loading a comma-separated list of multiple directories :

tensorboard --logdir_spec $(ls -m -d ./runs/2020-01-* | tr -d ' \n')

We use 2 ls flags here: -m sets delimiter to , (comma, space) and -d lists directories without their contents.

It assumes there are no commas or spaces in your log directory paths. You can use find for more complex cases.
More ways to concatenate list of paths with a comma

[wchargin]

@shtratos: The main point of this feature request is that the globs be
resolved dynamically by TensorBoard, such that if you were to add a new
./runs/2020-01-* directory at runtime it would show up without having
to restart the process.

If you’re satisfied with invocation-time glob resolution, you can use
the multitb shell function from the following comment, which is safer,
does not have restrictions on the characters in your path name, does not
use --logdir_spec, and works even if you have runs of the same name in
different sub-logdirs:
#179 (comment)

[LarsHill]

Any updates on this in 2022?

[roj4s]

You can create symbolic links of the selected list of experiments you want to upload. Let say you have a folder with experiments A, B, C and D, create a folder E and fill it with symbolic links to A and B only .. then run tensorboard dev upload on folder E.