More information for teardown in LightningDataModule

[yipliu]

Hi,

I found that a new function teardown has been added in MNISTDataModule, but all the descriptions of it on the web say to clean up the data.

Does it clean up the data on the memory? If so, it is very useful when we are working with very large data sets.

How do we use this function?

def teardown(self, stage: Optional[str] = None):
    """Clean up after fit or test."""
    pass

Don't write anything?

[ashleve]

According to lightning docs:
https://pytorch-lightning.readthedocs.io/en/latest/data/datamodule.html#teardown

The teardown is just a handy function called at the end of fit, validate, test, or predict.

It's up to you to decide what the function should be doing.

It is called from every process across all the nodes.

Thestage arg informs us whether the current stage is "fit", "validate", etc.

[yipliu]

Get it.

In general, what do we do with this function？

I hardly see any examples of using this function.

[ashleve]

• cleanning up downloaded data after training finishes
• closing any open connections a dataloader makes
• etc.

[nodiper]

def teardown(self, stage: Optional[str] = None):
"""Clean up after fit or test."""
pass