Pytorch-lightning: Nomenclature: reload dataloaders every epoch

[pietrolesci]

Simple nomenclature fix:

Since the trainer flag reload_dataloaders_every_epoch reloads only the training dataloader, as opposed to validation and training dataloaders (as implemented here), wouldn't it be better to change the nomenclature to reload_dataloader_every_epoch?

[github-actions[bot]]

Hi! thanks for your contribution!, great first issue!

[awaelchli]

Hi
it also reloads the other dataloaders.
See in your link a few lines down:
https://github.com/PyTorchLightning/pytorch-lightning/blob/f63fec9323c319720d78b452e9fe84b97ce7644e/pytorch_lightning/trainer/training_loop.py#L264

[pietrolesci]

HI @awaelchli,

Thanks a lot for the clarification. Is there a way I can force it to reload only the training set?

[awaelchli]

The only way I currently see is this:

# 1. Set reload_dataloaders_every_epoch=True
# 2. Store a reference to your val_dataloader in LightningModule and return it

def train_dataloader(self):
    # init train dataloader as usual
    ... 
    return train_dataloader

def val_dataloader(self):
    # will return a reference instead of recreating dataset

    if self.val_dataloader_ref:
       return self.val_dataloader_ref

    # init dataloader as usual
    ...
    self.val_dataloader_ref = val_dataloader
    return val_dataloader

[pietrolesci]

Great! Thanks a lot for the idea!
As far as this issue is concerned, I would suggest closing it

[awaelchli]

you're welcome and let me know if you run into further problems/questions.

[pietrolesci]

Thanks for your consideration @awaelchli!