Create a doc page about "Machine Learning on a 2i2c Hub"

[GeorgianaElena]

Context

We should document the GPU usage with pytorch, tensorflow, and jax, and talk about tensorboard.

Reference https://2i2c.freshdesk.com/a/tickets/234

Proposal

No response

Updates and actions

• refine this issue
• create a place or decide on a place for this documentation
• ...

[consideRatio]

The support ticket started out with @rabernat raising a question from another user.

Question for the jupyterhub users: do you use tensorboard? It is installed in the default environment, but I don’t understand how to launch the interface

I think ideally, this documentation should be in the tensorboard project and not managed by 2i2c. If 2i2c are to help here in a image specific question, I'd like see 2i2c help by helping tensorboard rather than providing our own island of assistance - which realistically won't be kept updated as fast as the tensorboard project evolves.

I think I lack some insight into past decisions in 2i2c about the level of support we aim to provide for software run in the user environments though.

[choldgraf]

A few thoughts from me:

I think this is a tension we've always felt in our documentation. How much do we want to rely on upstream (given that upstream docs are often for a more general audience, or go into more detail than we want) vs. centralize on our own docs (and thus incur extra documentation maintenance + centralize more attention in 2i2c spaces that should instead go to upstream communities). I'm still not sure what kind of balance is the best.

Here's a quick thought on what we could focus on:

• What is possible with our infrastructure. People need to know they can do something with our hubs before they can learn how to do it.
• Basic workflows and how the infrastructure works, but without going into too much detail and linking heavily to upstream docs.
• Conceptual guides that help people make decisions and understand what they want to do
• Anything that is unique to the combination of tools we expose (e.g. S3 buckets + a particular library).

I think Pangeo has a similar challenge + lots of documentation like this (e.g. their data guide). I believe that having the docs in that centralized space probably means a lot more people paid attention to them than if they would have had to go digging through the docs of a bunch of upstream packages.

cc also @jmunroe who I know has thought about this as well