Exact behaviour of .prepare_data() and .setup() in LightningDataModule when passing DM or dataloaders to Trainer in respect to internal hooks.

[tiefenthaler]

What is the exact behaviour of .prepare_data() and .setup() in LightningDataModule when passing DM or dataloaders to Trainer.
Which internal hooks are getting triggered in both scenarios?

What is given

Let´s assume the following is implemented:

• LightningDataModule (with .prepare_data() and .setup() and train, val, test, predict dataloaders)
• LightningModule (with train, val, test, predict steps)
• torch.nn.Module
• Trainer

Lightning Docs - Hooks:

• .prepare_data
  • This is called before requesting the dataloaders: model.prepare_data() -> initialize_distributed() -> model.setup(stage) -> model.train_dataloader() -> model.val_dataloader() -> model.test_dataloader() -> model.predict_dataloader()
• .setup
  • This is called before requesting the dataloaders
  • Called at the beginning of fit (train + validate), validate, test, or predict.
• Trainer.fit/val/test/predict
  • datamodule (Optional[LightningDataModule]) – A LightningDataModule that defines the train_dataloader hook.

When passing dataloaders to Trainer

• manual call of dm.prepare_data() and dm.setup(stage='fit') to get information from the dataset to build the model.
• trainer.fit(model=lightningmodel, train_dataloaders=dm.train_dataloader(), val_dataloaders=dm.val_dataloader())

1. Question: Based on docs, does this mean that .prepare_data and .setup are called two times respectively for dm.train_dataloader() and dm.val_dataloader()? Since two dataloaders are requested separately.
2. Question: Based on docs, does this mean that regardless if .prepare_data or .prepare_data+.setup were executed manually, they will be executed again (a 2. and 3. time), since two dataloaders are requested separately?
3. When calling Trainer.fit() it also triggers the .prepare_data and .setup hooks?
4. When calling Trainer.test(dataloaders=dm.test_dataloader()), .prepare_data and .setup are called again?
5. Is it possible to avoid the execution of .prepare_data and .setup in hooks?
6. I guess there must be some internal logic to handle this properly?

When passing LightningDataModule to Trainer

• manual call of dm.prepare_data() and dm.setup(stage='fit') to get information from the dataset to build the model.
• trainer.fit(datamodule=dm)

1. Question are the respective dataloaders be generated at once or separat (as above)?
2. Question: Based on docs, does this mean that regardless if .prepare_data or .prepare_data+.setup were executed manually, they will be executed again, since dataloaders are requested?
3. When calling Trainer.fit() it also triggers the .prepare_data and .setup hooks?
4. When calling Trainer.test(datamodule=dm), .prepare_data and .setup are called again?
5. 5. Is it possible to avoid the execution of .prepare_data and .setup in hooks?