Top hit for Pytorch on Ray isn't helpful for starting out

[worldveil]

Searching for "ray pytorch" leads to this page

We should update this page to be more beginner resources / an index, and move advanced best practices (current content) to another page via link.

[amogkam]

I think we can just remove this page altogether. It's pretty much all completely moot now with Ray Train. Just the "Downloading Data" Section we can move over to the Ray Train user guide. cc @maxpumperla @richardliaw

[worldveil]

Not sure I agree. We need a page that is the definitive place people land when they search "pytorch ray python". What will the contents be?

[richardliaw]

@maxpumperla to comment, but how about let's just wait a bit to address this? Things should make more sense with Ray ML higher-level structure.

[amogkam]

@worldveil thats just an SEO question right? When people search up "python ray pytorch" it should lead to the torch section of Ray ML.

[maxpumperla]

@ericl mentioned in Slack that we might want to move or remove some of the examples currently under Core. I think these 3 best practices guides (TF, Pytorch, notebooks) should be revisited.

We currently have good content for Pytorch for Tune and Train, and those should be surfaced in search first. Both are good destinations.

I think the page in question is interesting, because you learn how to use Ray Core to parallelise training. But then again, it's rather lengthy and I agree with @amogkam that we have train now and don't really need it. That's even more true since the page is called "best practice", which it certainly isn't. It might be instructive to have it for comparison in train, to show people how much work you save with train / all the stuff train does for you.

[worldveil]

I see in "master" this page is still around.

Can we have a link at top of page at least to direct people to an easier integration with PyTorch? it's not a good page for first time googlers of Ray + PyTorch...

[ericl]

Yeah, we can direct users with a big warning at the top. We did this before for RLlib (A3C has a info tip at the top pointing to RLlib: https://docs.ray.io/en/latest/auto_examples/plot_example-a3c.html).

It definitely definitely should not say this is "best practice", more like "example of using Ray to parallelise ML" as noted.

[maxpumperla]

@worldveil @ericl will take care of this, got enough info from Eric now

[maxpumperla]

x-posting from slack, following advise from @ericl:

Got it, here's my set of recommendations:

• Best practices for TF / Torch -> Remove or at least rename / redirect.
• Placement group -> Remove. Many of the examples seem misleading, and I think it overall does more harm than good.
• Streaming MR, News reader, Progress bar -> Remove. These are low quality and not mapping to real use cases.
• A3C, Model selection, Param Server, L-bfgs, Pong > Keep, but clarify Tune/RLlib/Train/etc should be used for real use cases.
• Tips -> Remove. These are redundant with Ray patterns already. Dev tips can be moved into contributors section.
• Pytorch lightning / XGBoost stuff -> Move into Ray ML section.

[maxpumperla]

Everything addressed in the PR linked above, we also have redirect rules in place for all deleted examples:

In particular, @worldveil 's search result should now redirect to Train examples, which have proper TF examples.

[maxpumperla]

@worldveil the file is now deleted, and at least in our algolia search on our docs the right documents are now surfaced when searching for pytorch. closing for now