Why does logic in processed_file_names make InMemoryDataset forget about processed_paths?

[NiklasKappel]

Consider this minimal (not) working example:

# file my_dataset.py

from torch_geometric.data import Data, InMemoryDataset


class MyOwnDataset(InMemoryDataset):
    def __init__(
        self,
        root,
        transform=None,
        pre_transform=None,
        pre_filter=None,
        remove_outliers=True,
    ):
        super().__init__(root, transform, pre_transform, pre_filter)
        self.remove_outliers = remove_outliers
        self.load(self.processed_paths[0])

    @property
    def raw_file_names(self):
        return ["data.txt"]

    @property
    def processed_file_names(self):
        if self.remove_outliers:
            return ["data_wo_outliers.pt"]
        return ["data.pt"]

    def process(self):
        data_list = []

        # Read data file
        data_list.append(Data())

        if self.remove_outliers:
            # Remove outliers
            pass

        self.save(data_list, self.processed_paths[0])


if __name__ == "__main__":
    dataset = MyOwnDataset(root="data")
    print(dataset[0])

Running python my_dataset.py crashes with:
AttributeError: 'MyOwnDataset' object has no attribute 'processed_paths'. Did you mean: 'processed_dir'?

If I comment out only the if condition in processed_file_names, the code works (assuming there is a file at data/raw/data.txt). I don't get how the presence of an if condition one method (processed_file_names) can determine the existence of a completely different method (processed_paths) in the parent class. I am not sure if this is a bug with PyG, or if I am doing something wrong here.

Edit: I can also make it work by keeping the if condition, but swapping the first to lines of __init__, that is the super call and the remove_outliers assignment. I have don't get why though, which makes me feel pretty uncomfortable with my solution.

[rusty1s]

Thanks for the issue. Python will output this error because there is an error raised when accessing processed_file_names. In particular, because self.remove_outliers is not yet defined since processed_file_names is called in super().__init__().

This fixes it:

        self.remove_outliers = remove_outliers
        super().__init__(root, transform, pre_transform, pre_filter)
        self.load(self.processed_paths[0])