Customize multiple dataloaders behaviour in train vs val/test

[edenlightning]

User should be able to determine if multiple dataloaders will run simultaneously (passed as a list/dict), or get called sequentially.
This can be done with a trainer flag with multiple options.

Background:
Currently, in training step you get the different batches at the same time as list or dict, but for validation step it gets called multiple times with a dataloader_idx. Just thought it would be nicer if how multiple dataloaders are handled is consistent across training, val and test.

[mees]

@williamFalcon's suggestion was to have something like Trainer(multi_dataset_style=“sequential | simultaneous”) to enable consistent multiple dataloader behaviors.

[tchaton]

Hey @mees @edenlightning,

Any consensus on the feature to be implemented ?

Note: Sequential training would be quick a large refactor internally to enable this feature.

Best,
T.C

[stale]

This issue has been automatically marked as stale because it hasn't had any recent activity. This issue will be closed in 7 days if no further activity occurs. Thank you for your contributions, Pytorch Lightning Team!