Passing transforms to datamodule from LightningCLI example

[cgoliver]

Would appreciate if someone has an example of passing a list of transforms with arguments to a Lightning DataModule from the LightningCLI.

I'm thinking of a syntax similar to this where one can pass multiple callables from the CLI.

EDIT: I managed to pass a single transform by adding the transforms: Optional[Callable] = None, typing hint to my datamodule __init__() but I get an unrecognized argument error when using the data.transforms+=my.transforms.Transform syntax in the CLI.

Thanks!

[tshu-w]

You should use list, for more details, see https://jsonargparse.readthedocs.io/en/stable/#type-hints

[SinaSarparast]

I'm also trying to initiate transformations and pass them to the data module. I use the following config for the data module:

data:
  class_path: pfp.data.data_loader.DataLoaderIter
  init_args:
    batch_size: 2
    pyg_file_dir: ./pyg_dataset
    transformations:
      - class_path: torch_geometric.transforms.center.Center

When debugging this is parsed as :

transformations = [{'class_path': 'torch_geometric.tran...ter.Center'}]

The object is not initiated like the callbacks that are passed to the trainer. Is this how it should be or I'm doing something wrong?

[tshu-w]

#15115