LightningCLI: Passing objects via link_arguments

[lneukom]

Hi,

is there a way to pass objects from a LightningDataModule to a LightningModule in the LightningCLI using link_arguments? I tried the following

class TestModel(LightningModule):

    def __init__(self, scaler: StandardScaler, *args: Any, **kwargs: Any):
        super().__init__(*args, **kwargs)
        print(scaler)


class TestDataModule(LightningDataModule):

    def __init__(self):
        super().__init__()
        self.scaler = StandardScaler()


class TestCLI(LightningCLI):

    def add_arguments_to_parser(self, parser):
        parser.link_arguments('data.scaler', 'model.scaler', apply_on='instantiate')


TestCLI(TestModel, TestDataModule)

but got a ValueError: Target key "model.scaler" must be for an individual argument

[lneukom]

@mauvilsa posted here as discussed in slack

[mauvilsa]

This is a valid use case for link_arguments but currently not supported. I will look at what needs to be done to make this work and comment here.

[mauvilsa]

The support for this has been implemented in jsonargparse. There is no need for any change in pytorch-lightning. Just update the package, i.e. pip install -U jsonargparse, and your example should work.

[mauvilsa]

@lneukom do let me know if you tried it and if it worked as you expected.

[lneukom]

Just tested it, works! Thanks for the quick fix!