I bundled up a quick proof of concept for having tensorboard outside tensorflow.
It utterly cheats because it actually is invoking TensorFlow.jl to do this.
It would be possible at some point to build something similar (or modify this),
to directly speak tensorboard using something like https://github.com/TeamHG-Memex/tensorboard_logger
https://github.com/oxinabox/UniversalTensorBoard.jl
If you follow the examples in the readme, it should be pretty clear how you would get e.g. Flux to work with this via it’s callbacks.