create import protection for torchvision

[felipemello1]

in utils, it imports torchvision even if the recipe doesnt need it. We should protect this import, in case the user hasnt installed it and doesnt need it.

[joecummings]

What's your implementation plan @felipemello1 ? I'd prefer not to add a tag "good first issue" without really clear instructions.

[felipemello1]

This is the error:

  File "/data/users/felipemello/torchtune/torchtune/data/__init__.py", line 7, in <module>
    from torchtune.data._chat_formats import (
  File "/data/users/felipemello/torchtune/torchtune/data/_chat_formats.py", line 10, in <module>
    from torchtune.data._messages import Message, Role
  File "/data/users/felipemello/torchtune/torchtune/data/_messages.py", line 9, in <module>
    from torchtune.modules.transforms import Transform
  File "/data/users/felipemello/torchtune/torchtune/modules/transforms/__init__.py", line 17, in <module>
    from torchtune.modules.transforms.vision_utils.resize_with_pad import (  # noqa
  File "/data/users/felipemello/torchtune/torchtune/modules/transforms/vision_utils/resize_with_pad.py", line 14, in <module>
    import torchvision
ModuleNotFoundError: No module named 'torchvision'

It happens because in transforms.init we expose vision transforms.

A simple solution is to not import chat transforms from the same place as vision transforms. This can be done by renaming /vision_utils to /vision_transforms, and importing from there.

And putting the transforms.py (chat related) under /text_transforms, and importing from there.

cc: @RdoubleA @pbontrager

[joecummings]

I agree this needs to be better thought out how we structure our imports here. However another stopgap we could do is the following:

try:
	import torchvision
except ImportError:
	log.warning("Attempted to import torchvision. Feel free to ignore if you are not trying to do multimodal training. Otherwise, please install ``torchvision`` package via conda or pip")

This would immediately allow users to use our package without the pre-requisite of another heavy PyTorch library.

I acknowledge it's hackiness, but in my mind this is a bad starting experience and it's not immediately clear ATM why someone would need torchvision when we don't (yet) support MM training.

Thoughts @felipemello1 @pbontrager ?

[felipemello1]

wouldnt you have to add it in all files that import torchvision? If you want a quick solution, I think that moving these functions from transforms/init --> transforms/vision_transforms/init will be best.

[joecummings]

wouldnt you have to add it in all files that import torchvision? If you want a quick solution, I think that moving these functions from transforms/init --> transforms/vision_transforms/init will be best.

Fair enough - I still consider that a stopgap, but I see your point that it's less intrusive than a try/except.