'tensorboard.plugins.projector' has no attribute 'ProjectorConfig'

[marcevrard]

Referring to the test file:
https://github.com/tensorflow/tensorboard/blob/1.5/tensorboard/plugins/projector/projector_api_test.py

If projector is imported as:

from tensorboard.plugins import projector
config = projector.ProjectorConfig()

accessing the ProjectorConfig() raises:
AttributeError: module 'tensorboard.plugins.projector' has no attribute 'ProjectorConfig'

It works fine if projector is imported from tf.contrib:

from tensorflow.contrib.tensorboard.plugins import projector

---

tensorboard == 1.5.0
tensorflow == 1.5.0
Python == 3.6.3
Ubuntu == 17.10

[chihuahua]

Hmm, importing ProjectorConfig should work due to this line:

tensorboard/tensorboard/plugins/projector/__init__.py

Line 35 in f6bb9b0

from tensorboard.plugins.projector.projector_config_pb2 import ProjectorConfig

I wonder why not. I wonder what happens if you install tensorboard nightly (pip install tb-nightly) and thus run on the bleeding edge version.

[marcevrard]

Just tried and got the exact same error.

[nfelt]

Upon further investigation, it looks like the actual contents of our __init__.py files isn't being preserved during our pip package building. We didn't notice because most of them are empty (aside from a license) and Bazel evidently creates its own empty __init__.py files by default: bazelbuild/bazel#4470

The projector plugin is pretty much the only non-empty one aside from the top-level one, which is specifically listed in its Bazel rule and does get included. I think if we start explicitly listing the other ones that should fix this issue.

[nfelt]

For now, a workaround should be to explicitly import from projector_config_pb2 as in

from tensorboard.plugins.projector.projector_config_pb2 import ProjectorConfig
config = ProjectorConfig()

[marcevrard]

Though the function visualize_embeddings() won't be available either.
My current workaround is thus to import from:

from tensorflow.contrib.tensorboard.plugins import projector

Are there any possible issues using the tf.contrib module instead of the tb root module? (I guess this will be still supported for a while)

[nfelt]

Oh, yes, you're right. I believe that will work for now but it will be safer to import from tensorboard once we get that working.

[marcevrard]

I checked with the latest package (tb-nightly-1.6.0a20180206) and it works perfectly.
Thanks for the swift fix!

[nfelt]

Sure. This should also be fixed in 1.5.1 (just released yesterday) and will be in the upcoming 1.6.0 as well.