Rosario Scalise

If this isn't likely to be supported anytime soon, does anyone have an appropriate work around? I'm also stuck on this.

Nice! Thanks @graingert! It looks like @adamchainz (OP) got tired of the problem and created this shortly after filing the issue 😆 . For the record, I ended up adding...

Hi @ashleve this is all very relevant to the efforts @omry and I have been making. We would love to have your input and help! Easiest way to chat about...

Sorry about that my fellow rosario! Hahaha

Before merging, I just want to point out that because of the way pytorch structures its files (and `__init__.py` for certain modules), the optimizer register can be done like this:...

@briankosw If you have a draft PR, I'm happy to check it out! =] Doesn't need to be finished.

> > High level feedback: > > We will probably have multiple examples for distributed data parallel, with different limitations and advantages. > > It's good to group them together...

Don't the users have control by determining which 'module' they import? What was your alternative proposition?

Ok, I think this is probably a better option. The automatic strategy would likely leave users confused about what _is_ and what _isnt_ currently registered. Better to be explicit and...

I think there are at least 3 meaningful examples to work on: 1.) Barebones, configuring just one thing (like the optimizer) 2.) Full MNIST configuration. 3.) Training framework like Lightning...