Extract dataset definition out of the LightningModule

[cgarciae]

🚀 Feature

Extract dataset definition out of the LightningModule

Motivation

Separation of data from the model.

Pitch

The datasets loaders could easily be passed to the fit method directly instead of having to define them inside the LightningModule, this avoids having a single class that possibly contains: data, data pipeline params, model, model hyperparams.

The basic example cloud look like this:

from pytorch_lightning import Trainer

train_dataloader = DataLoader(...)
val_dataloader = DataLoader(...)

model = CoolSystem()

# most basic trainer, uses good defaults
trainer = Trainer()    
trainer.fit(model, train_dataloader=train_dataloader, val_dataloader=val_dataloader)

Its much more natural to how you usually structure your code in scikit-learn or keras.

Alternatives

They could also be based to the Trainers constructor.

[williamFalcon]

Why not allow for both? add 3 args to trainer or do it in .fit().

if any of these datasets are passed in the corresponding lightningModule class isn’t called.

thoughts?

[williamFalcon]

@cgarciae want to submit this PR? i agree that we should allow both ways of doing it

[awaelchli]

@cgarciae can the issue be closed or is there something left to do?

[cgarciae]

Looks awesome!

[williamFalcon]

@lorenzoFabbri @tullie @myleott @ashwinb @shootingsoul @vreis @Darktex

FYI