DataModules: skip prepare_data

[adamjstewart]

Summary

We should skip the prepare_data step in our data modules unless downloading is actually required.

Rationale

Instantiating the dataset is slow. Currently, all of our datamodules instantiate the dataset at least twice on the off chance that someone is requesting the dataset to be downloaded. This isn't needed most of the time.

Implementation

A better solution would be to replace:

def prepare_data(self) -> None:
    Dataset(**self.kwargs)

with this:

def prepare_data(self) -> None:
    if self.kwargs.get("download", False):
        Dataset(**self.kwargs)

in all of our data modules.

Alternatives

No response

Additional information

No response

[iamhbc]

The solution is to replace the prepare_data() method with the code above in all of our data modules. This code will check the "download" keyword argument and only download the dataset if it is set to true. This will significantly reduce the time required to instantiate the dataset and speed up our data modules.