`pip install lightning-flash[image]` results in torch downgrade to 1.8.1

🐛 Bug

One of the dependencies of lightning-flash requires torch version <= 1.7.1, this is untenable for my use case and really needs to be fixed. I believe it’s an issue with icevision. As you can see icevision requires torch version strictly less than 1.9.

To Reproduce

Have torch > 1.9 installed in a conda environment. Attempt to install lightning-flash[image]. Obeserve at some point pip attempting to downgrade torch. This has now occurred on two separate machines with separate architectures, which isn’t a huge issue on Linux (my requirement of 1.9 is on Windows), but on Windows it replaces the GPU version of pytorch with the CPU version.

Environment

• PyTorch Version (e.g., 1.0): 1.9
• OS (e.g., Linux): Linux + Windows
• How you installed PyTorch (conda, pip, source): conda
• Build command you used (if compiling from source):
• Python version: 3.7 and 3.8

Additional context

I can trick it a bit if I install it with no dependencies. This is what I’ve had to do and then manually install each missing package as it comes up.

[ethanwharris]

Hi @dlangerm whoops, meant to close this issue. IceVision is now a fully optional dependency in flash 0.5.0 😃 only required for the IceVision based tasks (object detection, keypoint detection, and instance segmentation).

[ethanwharris]

Hi @dlangerm thanks for reporting this! Yeah, this seems to be caused by the install of icevision. I think we can do a small refactor here to make this work, just move icevision to a fully optional dependency for a handful of tasks.