Allow Using Specific Dataloaders in LightningCLI (Instead of a DataModule)

[stuartthomson]

Description & Motivation

I would like to configure the LightningCLI to use dataloaders in the test subcommand, similar to this example from the docs

trainer.test(dataloaders=test_dataloaders)

As far as I can tell this impossible using the LightningCLI because you can only pass a datamodule into it. My test set is defined differently from various training sets I have so this becomes annoying to create separate datamodules for each.

You can see that the dataloaders option is specifically ignored in the cli.py subcommands method.

Pitch

I would like to specify dataloaders in the LightningCLI. Perhaps like:

LightningCLI(
  MyModule,
  train_dataloaders=MyTrainDataloader,
  test_dataloaders=MyTestDataloader,
  args=args
)

Alternatives

The alternative is having to create different datamodules for every combination of train/test dataloaders.
Or you have to simply just not use the CLI stuff and call lightning directly.

Additional context

I would be willing to open a PR if this feature is deemed useful.

cc @lantiga @Borda

[stale]

This issue has been automatically marked as stale because it hasn't had any recent activity. This issue will be closed in 7 days if no further activity occurs. Thank you for your contributions - the Lightning Team!