Support CLI shorthand natively

[mauvilsa]

What does this PR do?

Uses a new version of jsonargparse which implements for subclasses the support for short command line arguments and specifying only the class name instead of the full class path. This allows to remove the placeholder LightningArgumentParser._convert_argv_issue_84.

The idea with the 4.6.0 version of jsonargparse is that all subclasses for the given base classes can be specified by name. No need to explicitly register them. The only requirement is that the module where the subclass is defined, is imported prior to parsing. For example in the case of --lr_scheduler two base classes are given torch.optim.lr_scheduler.{_LRScheduler,ReduceLROnPlateau}. All subclasses in the same module are already known by the parser, so for instance --lr_scheduler ExponentialLR would already work. One nice benefit of this is that many other parameters get the same support without needing to implement a register for it, e.g. --trainer.accelerator. And also works for user defined parameters, e.g. a module with signature def __init__(self, encoder: EncoderBase, decoder: DecoderBase):.

This does not introduce any breaking change, though it does make the registries a bit misleading. The CLIs might support more classes than what the registry has. Also using a decorator, e.g. @MODEL_REGISTRY, makes no difference. The module where a class is defined has to be imported for the decorator to register it. But since it is imported, the parser already knows about this class and supports it without having to register it.

Until _convert_argv_issue_85 is removed, still there is a need to register the callbacks. But assuming that this is done, there is a general question. Should the registries be deprecated?

Fixes #12464

Does your PR introduce any breaking changes? If yes, please list them.

None

Before submitting

• Was this discussed/approved via a GitHub issue? (not for typos and docs)
• Did you read the contributor guideline, Pull Request section?
• Did you make sure your PR does only one thing, instead of bundling different changes together?
• [n/a] Did you make sure to update the documentation with your changes? (if necessary)
• Did you write any new necessary tests? (not for typos and docs)
• Did you verify new and existing tests pass locally with your changes?
• Did you list all the breaking changes introduced by this pull request?
• Did you update the CHANGELOG? (not for typos, docs, test updates, or internal minor changes/refactorings)

PR review

Anyone in the community is welcome to review the PR.
Before you start reviewing make sure you have read Review guidelines. In short, see the following bullet-list:

• Is this pull request ready for review? (if not, please submit in draft mode)
• Check that all items from Before submitting are resolved
• Make sure the title is self-explanatory and the description concisely explains the PR
• Add labels and milestones (and optionally projects) to the PR so it can be classified

Did you have fun?

Make sure you had fun coding 🙃

[carmocca]

Until _convert_argv_issue_85 is removed, still there is a need to register the callbacks. But assuming that this is done, there is a general question. Should the registries be deprecated?

If we have no use for them, yes.

[carmocca]

This is great!

Follow-ups would be:

• Deprecate registries
• Deprecate LightningCLI(auto_registry=...)
• Rewrite docs

[mauvilsa]

This is missing the milestone. I guess it would be 1.7, right?

[rohitgr7]

hey @mauvilsa!
mind check the failing tests?

[mauvilsa]

hey @mauvilsa! mind check the failing tests?

This is why I implemented mock_subclasses, see https://github.com/mauvilsa/pytorch-lightning/blob/cli-remove-hack-84/tests/utilities/test_cli.py#L65-L73. Should be fixed by just adding the mock in that test.