load_from_checkpoint support for LightningCLI when using dependency injection

[mauvilsa]

What does this PR do?

For now this pull request is intended to start the discussion on how to add full support for load_from_checkpoint for models trained using LightningCLI.

LightningCLI is designed to allow the use of dependency injection, which is a good programming pattern. Since this means that class instances are given to the module's __init__, the current implementation of save_hyperparameters is not enough because there is no reliable way to know how the objects were instantiated. And this prevents load_from_checkpoint from working correctly.

The main idea is that LightningCLI provides to save_hyperparameters, through a context variable, the actual parameters used for instantiation. These get saved as normal in the hparams.yaml file. Then load_from_checkpoint uses a custom instantiator that makes use of jsonargparse to support the configuration of subclasses (class_path and init_args) in configs.

The issue #15427 asks for a way to instantiate a model from a config file. However, load_from_checkpoint does both instantiation and loading of weights. Is there really a need to only instantiate from config? Like having a load_from_config which would get the path to the hparams.yaml file?

Also #15427 asks about dataloaders. From what I understand save_hyperparameters can also be used for data modules. However, I don't know how that is intended to work.

Fixes #15427
Fixes #13279

Before submitting

• Was this discussed/agreed 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?
• 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 minor internal changes/refactors)

PR review

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

Reviewer checklist

• 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

[Borda]

Can we have this as optional so that still allow users to use an older version of jsonargparse and not have a too narrow version range...

[carmocca]

Interesting solution. But I think we should iterate on the UX for this first

[carmocca]

Overall looks fine. Given the complexity and the lack of docs I don't expect users to try this yet. Will you add those later?

[codecov]

Codecov Report

Merging #18105 (c71406f) into master (39a86f8) will decrease coverage by 35%.
Report is 3 commits behind head on master.
The diff coverage is 100%.

Additional details and impacted files

@@            Coverage Diff             @@
##           master   #18105      +/-   ##
==========================================
- Coverage      84%      48%     -35%     
==========================================
  Files         450      442       -8     
  Lines       38154    38050     -104     
==========================================
- Hits        31893    18430   -13463     
- Misses       6261    19620   +13359     

[carmocca]

@Borda you have changes requested for this PR. Can you check it again please?