predict_dataloader is missing in LightningDataModule

[digital-idiot]

In the documentation it is said that:

Use the predict_dataloader() method to generate the prediction dataloader(s)

However, unlike train_dataloader, val_dataloader and test_dataloader there is no reference of predict_dataloader in the source. Also from_dataset class method does not have any argument for passing predict_dataset.

[rohitgr7]

hey @digital-idiot !
good catch! would you like to send a PR to add it?

[digital-idiot]

@rohitgr7
Yes, I may, but it may take a day or two.

[digital-idiot]

@rohitgr7
When I was trying use LightningDataModule I also noticed that train_dataloader is using shuffle=True by default which user can't change also for my personal project I needed to provide collate_fn which I can't do ATM without overloading the method. Would it be acceptable if the API just provides a general **kwargs option where we can pass all the options supported by DataLoader?

[rohitgr7]

tbh, I was going to propose to deprecate and remove this method 😅 since it just solved 2 lines of code and doesn't let users to follow the general approach of using prepare_data and setup.

also adding **kwargs would make it difficult to configure them since we can't say whether they should be added for all dataloaders or just any one of them. So I'd always recommend using setup and prepare_data and override dataloader methods to define the dataloader.