AttributeError: BinaryMNISTDataModule object has no attribute train_transforms

[davidgilbertson]

🐛 Bug

To Reproduce

Trying to run the code from the docs here: https://lightning-bolts.readthedocs.io/en/stable/datamodules/vision.html#binarymnist

Steps to reproduce the behavior: Run the code from that page:

from pl_bolts.datamodules import BinaryMNISTDataModule

dm = BinaryMNISTDataModule('.')
model = LitModel()

Trainer().fit(model, datamodule=dm)

Expected behavior

It shouldn't error :)

Environment

pytorch-lightning: 1.7.7
lightning-bolts: 0.5.0

Additional context

Adding this makes it work:

dm = BinaryMNISTDataModule(".")
dm.train_transforms = None
dm.val_transforms = None
...

This check (and the val counterpart) are assuming the existence of an attribute that isn't defined as a default.

train_transforms = self.default_transforms() if self.train_transforms is None else self.train_transforms

From https://github.com/Lightning-AI/lightning-bolts/blob/master/pl_bolts/datamodules/vision_datamodule.py#L103

I don't know the codebase well, but it looks like either train_transforms should have a default value, or a hasattr() check should be added, or query the value with a fallback, like if getattr(self, 'train_transforms', None) is None

Or maybe I'm doing something wrong. The docs have this cryptic little message about transforms, with no context. What is this saying, what is transform_lib, what am I supposed to do with mnist_transforms?

[otaj]

Hi, @davidgilbertson. This is happening because pytorch-lightning>=1.7.0 removed these methods and lightning-bolts==0.5.0 was released quite some time ago.

If you update to lightning-bolts==0.6.0, this should be fixed. We made sure we're being a bit more compatible with current version.

OTOH, beware please, that if you use 0.6.0, you will get quite a bit of warnings originating from our package, this is mainly due to issue #839. They should be harmless, but there will be quite many of them