What is the on_train_dataloader hook expected to do?

[ananthsub]

https://github.com/PyTorchLightning/pytorch-lightning/blob/963c26764682fa4cf64c93c5a7572ae0040e9c32/pytorch_lightning/core/hooks.py#L669-L670

Why do we have these hooks? What does it mean to “request” these dataloaders? What logic goes here vs inside the train/val/test/predict dataloader methods directly? What guarantees does the trainer make about when or how often these hooks are called?

[carmocca]

Why do we have these hooks?

cc @tchaton who added these in 0995d30

Seems like they are used to inject logic in Flash

https://github.com/PyTorchLightning/lightning-flash/blob/4ebb0877f3d406919fc07791369f0a33643c1ec5/flash/core/model.py#L504-L526

cc @ethanwharris

What does it mean to “request” these dataloaders?

To call model.*_dataloader() or datamodule.*_dataloader(), the verb request was probably used due to the name of the function that calls it:

https://github.com/PyTorchLightning/pytorch-lightning/blob/4928dc55790eec4ee4a558f494a6fc3d49a5e78a/pytorch_lightning/trainer/data_loading.py#L503

What guarantees does the trainer make about when or how often these hooks are called?

The only guarantee is that it will get called when the dataloader functions do.

[ananthsub]

@ethanwharris @tchaton - since it's called at the same time model.*_dataloader() or datamodule.*_dataloader() are called, can this logic sit directly in train_dataloader()? what justifies the separate hook?

[ethanwharris]

@ananthsub I'm not sure on the details, but it seems to me like the on_*_dataloader hooks are not needed and could be deprecated. At least, if we deprecate it and it turns out we really need it (or someone else does) then we can always undeprecate haha

[ananthsub]

filed #8758 to discuss further

[tchaton]

Hey @ananthsub,

Those hooks were used to perform in-place modification every time the data-loaders were to be created.
We could deprecate them as there is other alternative within Flash now.

Best,
T.C