Satlas pretrained Swin Transformer v2 models

[isaaccorley]

Adds Swin Transformer v2 Base pretrained models from the Satlas dataset

[adamjstewart]

Will review after deadline

[calebrob6]

is the timm swin transformer compatible?

[isaaccorley]

No it's not. The layers are named differently unfortunately.

[isaaccorley]

Wonder if I can map them to each other.

[adamjstewart]

If it's not too hard, that would be useful since it would let us use it in our trainers. Wonder if we could add some kind of logic to support both timm and torchvision in our trainers...

[isaaccorley]

It is hard unfortunately.

[isaaccorley]

I have no idea why the figure plotting in torchgeo/trainers/detection.py is somehow indirectly not covered via codecov

[adamjstewart]

Turns out the decrease in test coverage is actually very important!

If you remove the try-except from ObjectDetectionTask, you'll find that there is a bug in our implementation:

FAILED tests/trainers/test_detection.py::TestObjectDetectionTask::test_trainer[True-faster-rcnn-nasa_marine_debris] - ValueError: Boxes need to be in (xmin, ymin, xmax, ymax) format. Use torchvision.ops.box_convert to convert them
FAILED tests/trainers/test_detection.py::TestObjectDetectionTask::test_trainer[True-fcos-nasa_marine_debris] - ValueError: Boxes need to be in (xmin, ymin, xmax, ymax) format. Use torchvision.ops.box_convert to convert them
FAILED tests/trainers/test_detection.py::TestObjectDetectionTask::test_trainer[True-retinanet-nasa_marine_debris] - ValueError: Boxes need to be in (xmin, ymin, xmax, ymax) format. Use torchvision.ops.box_convert to convert them
FAILED tests/trainers/test_detection.py::TestObjectDetectionTask::test_no_rgb[True] - ValueError

These were being masked by our try-except. The code works fine for torchvision 0.13 but breaks for any newer version. We didn't notice this because the combined coverage didn't change. From what I can tell, this has always been broken...

[adamjstewart]

Can you try to fix that in a separate PR so we can backport it to 0.5.2? Or one of us can do it, just let me know if you're busy.

[isaaccorley]

Can you try to fix that in a separate PR so we can backport it to 0.5.2? Or one of us can do it, just let me know if you're busy.

I'll push a PR here shortly

[adamjstewart]

Can merge as soon as coverage issue is fixed.