Accessing DataLoaders

In the case that you require access to the or objects, DataLoaders for each step can be accessed via the trainer properties train_dataloader(), val_dataloaders(), test_dataloaders(), and predict_dataloaders().

dataloaders = trainer.train_dataloader
dataloaders = trainer.val_dataloaders
dataloaders = trainer.test_dataloaders
dataloaders = trainer.predict_dataloaders

These properties will match exactly what was returned in your *_dataloader hooks or passed to the Trainer, meaning that if you returned a dictionary of dataloaders, these will return a dictionary of dataloaders.

Replacing DataLoaders

If you are using a CombinedLoader. A flattened list of DataLoaders can be accessed by doing:

from lightning.pytorch.utilities import CombinedLoader

iterables = {"dl1": dl1, "dl2": dl2}
combined_loader = CombinedLoader(iterables)
# access the original iterables
assert combined_loader.iterables is iterables
# the `.flattened` property can be convenient
assert combined_loader.flattened == [dl1, dl2]
# for example, to do a simple loop
updated = []
for dl in combined_loader.flattened:
    new_dl = apply_some_transformation_to(dl)
# it also allows you to easily replace the dataloaders
combined_loader.flattened = updated