[DataModule] prepare_data() and setup() not called

[remisphere]

🐛 Bug

It seems that when using DataModule to separate training logic and data loading,
of the five methods that should be called that are
prepare_data(), setup(), train_dataloader(), val_dataloader() and test_dataloader(),
only the last three are actually used, witch is problematic since the datasets used by the data-loaders should be assigned in the setup().

To Reproduce

Steps to reproduce the behavior:
Run this:

Code sample

import torch
from pytorch_lightning import LightningDataModule
from pytorch_lightning.core.lightning import LightningModule
from pytorch_lightning.trainer import Trainer
from torch.nn import L1Loss, Linear
from torch.optim import SGD
from torch.utils.data import DataLoader


class MyDataModule(LightningDataModule):

    def __init__(self):
        super().__init__()

    def prepare_data(self):
        print('in prepare_data, '
              'this should be called before train_dataloader() but is not.')

    def setup(self, stage):
        print('in setup, '
              'this should be called before train_dataloader() but is not.')
        self.train_dataset = 'whatever'

    def train_dataloader(self):
        print('in train_dataloader')
        return DataLoader(self.train_dataset)


class MyLightningModule(LightningModule):

    def __init__(self):
        super().__init__()
        self.layer = Linear(1, 1)
        self.loss_function = L1Loss()

    def forward(self, x):
        return self.layer(x)

    def configure_optimizers(self):
        return SGD(self.parameters(), lr=0.01)

    def training_step(self, batch, batch_idx):
        print("you won't even get here")
        raise NotImplementedError


data_module = MyDataModule()
model = MyLightningModule()
trainer = Trainer(gpus=1)
trainer.fit(model, data_module)

this gives AttributeError: 'MyDataModule' object has no attribute 'train_dataset'.

Expected behavior

When entering train_dataloader(), prepare_data() and setup() should already have been executed, and thus the train_dataset attribute should exist.

Additional context

IMHO, it comes from here

Environment

• CUDA:
  • GPU:
    • GeForce RTX 2080 Ti
    • GeForce RTX 2080 Ti
  • available: True
  • version: 10.1
• Packages:
  • numpy: 1.19.1
  • pyTorch_debug: False
  • pyTorch_version: 1.5.1+cu101
  • pytorch-lightning: 0.9.0rc2
  • tensorboard: 2.3.0
  • tqdm: 4.48.0
• System:
  • OS: Linux
  • architecture:
    • 64bit
  • processor: x86_64
  • python: 3.7.6
  • version: Unable to import trainer #41-Ubuntu SMP Tue Dec 3 00:27:35 UTC 2019

[nateraw]

1. you're not specifying the datamodule kwarg in trainer.fit() - your last line should look like this: trainer.fit(model, datamodule=data_module)

2. In this first iteration of LightningDataModule, you have to call setup and prepare_data manually for the datamodule instance. We have it set up this way so if you don't want to use Lightning, you can use your datamodule's loaders with pure Pytorch. I thought of having them called implicitly in the PR, but ended up landing on this for now. I'm not sure if users would always want these to run implicitly.

TL;DR: you can update your code to look like this:

# Init a datamodule
dm = MyDataModule()

# Manually call prepare_data and setup. You could put this at end of __init__ if you want
dm.prepare_data()
dm.setup()

model = MyLightningModule()
trainer = Trainer(gpus=1)
trainer.fit(model, datamodule=dm)

---

That being said, we're open to any ideas on making this more intuitive, so feel free to throw out some alternatives. 😄

[remisphere]

1. is not true in 0.9.0rc2: a data module as second positional argument is taken care of here.

2. I don't have a global enough view to know what other users might want, so if it is a feature i'm fine with it.
   I just saw that the manual call was in the docs, my bad for not looking far enough.

Anyway thank you for the clear answer ^^

[nateraw]

@remisphere I totally didn't notice! You were completely right on the dm arg. things move fast haha.

Reopening actually, as I think your intended use is more user friendly.