Tensorflow offers a nice visualisation tool for the model parameters: TensorBoard | TensorFlow
It can be launched directly from a notebook (or from the CLI), but then one needs to access it on another port. Is there any way to make this work with SWAN?
Of course, the current alternative is just to launch the tensorboard locally using the synce’d cernbox folder.
Thank you for the suggestion. This would be very nice to have indeed, even more now that we are starting to offer GPUs and GPU-enabled tensorflow in SWAN.
Is it absolutely necessary that there is a port open to the outside? Ideally, the notebook frontend should proxy the tensorboard, which would be opened as a local process listening on a local port. But to the outside everything should flow via the Jupyter websockets connection.
I wonder if that’s the case for e.g. :
Since we are moving soon(ish) to JupyterLab, it would be better to go for a JLab extension.
Yes exactly, I tried installing on SWAN and the above screenshot is a result from that. I can’t see the jupyter-tensorflow in the notebook though…
Thanks a lot for checking!
Thanks for debugging Omar!
Yes I wonder how such extensions can be integrated in SWAN’s custom interface? Do you think it needs some further custom rework from your side then?
And: when is the JLab expected to go to production?
Artur
@alobanov if we go for one, we’ll go for the JLab one. But before, the extension needs to be migrated to JLab 3 (and fix the problems pointed out by Omar).