Allow nested `batch_arg_name` in `BatchSizeFinder`/`Tuner.scale_batch_size()`

[ibro45]

Description & Motivation

Hi,

Would it be possible to allow dot-notation in, for example, tuner.scale_batch_size(model, batch_arg_name="dataloaders.train.batch_size")?

Perhaps some other features would benefit from this. It should be simple to achieve that through lightning_hasattr and lightning_getattr.

Pitch

No response

Alternatives

No response

Additional context

No response

cc @lantiga @Borda

[ibro45]

I added this to my class Model(LightningModule) that allows for that behavior, but I don't think this is ideal.

    def __getattr__(self, name: str) -> Any:
        """
        Override __getattr__ to handle dot-delimited attribute access.

        Args:
            name (str): The attribute name, potentially dot-delimited.

        Returns:
            Any: The value of the nested attribute.

        Raises:
            AttributeError: If the attribute path does not exist.
        """
        if '.' in name:
            try:
                attrs = name.split('.')
                obj = self
                for attr in attrs:
                    obj = getattr(obj, attr)
                return obj
            except AttributeError as e:
                raise AttributeError(f"Attribute path '{name}' not found.") from e
        # If no dot, defer to the default behavior
        return super().__getattr__(name)
    
    def __setattr__(self, name: str, value: Any) -> None:
        """
        Override __setattr__ to handle dot-delimited attribute setting.

        Args:
            name (str): The attribute name, potentially dot-delimited.
            value (Any): The value to set.

        Raises:
            AttributeError: If the attribute path does not exist.
        """
        if '.' in name:
            attrs = name.split('.')
            obj = self
            for attr in attrs[:-1]:
                try:
                    obj = getattr(obj, attr)
                except AttributeError as e:
                    raise AttributeError(f"Attribute path '{'.'.join(attrs[:-1])}' not found.") from e
            setattr(obj, attrs[-1], value)
        else:
            # Use the default behavior for single attributes
            super().__setattr__(name, value)

[SkafteNicki]

Hi and thanks for raising this issue. I took a look at what it would take to implement this feature and sadly it would require some major changes. However, a good alternative in my mind is that you can use a batch_size property/setter to still achieve what you want:

class DataLoaderConfig:
    """Helper class to hold dataloader configuration."""
    
    def __init__(self, batch_size=32):
        self.batch_size = batch_size

class NestedModel(LightningModule):
    """Model with nested batch_size attribute - this is what the issue wants to support."""
    
    def __init__(self, batch_size=32):
        super().__init__()
        self.linear = torch.nn.Linear(10, 1)
        self.dataloaders = DataLoaderConfig(batch_size)  # Nested attribute
    
    def train_dataloader(self):
        dataset = create_dummy_dataset()
        return DataLoader(dataset, batch_size=self.dataloaders.batch_size, shuffle=True)

    @property
    def batch_size(self):
        return self.dataloaders.batch_size

    @batch_size.setter
    def batch_size(self, value):
        self.dataloaders.batch_size = value 

In this way the batch_size attribute will be present during the BatchSizeFinder process but the end result will still be saved in the user specified attribute.
Closing issue for now, but feel free to ping and reopen if necessary.