Add data module for LEVIR-CD+ dataset

[robmarkcole]

Close #1706

[robmarkcole]

Error on cicd: RuntimeError: The daily quota of the file LEVIR-CD+.zip is exceeded and it can't be downloaded. This is a limitation of Google Drive and can only be overcome by trying again later.

After updating the image type to float the plotting broke, now addressing that too

[adamjstewart]

No tests should require internet access, you should use the fake data in tests/data/levircd. You may have to extract the zip file and commit all individual files so that pytest doesn't create new files that aren't tracked by git.

[robmarkcole]

We format the mask as torch.long but loss calc requries float - why do we not format it as such?

[robmarkcole]

Test fails with

FAILED tests/datamodules/test_levircd.py::TestLEVIRCDPlusDataModule::test_val_dataloader[train] - 

torchgeo.datamodules.utils.MisconfigurationException: LEVIRCDPlusDataModule.val_dataset has length 0.

which is why the plot test fails also

[adamjstewart]

We format the mask as torch.long but loss calc requries float - why do we not format it as such?

Kornia augmentations require it to be torch.long. Really wish the Kornia and torchmetrics devs would talk to each other.

[adamjstewart]

Test fails with

FAILED tests/datamodules/test_levircd.py::TestLEVIRCDPlusDataModule::test_val_dataloader[train] - 

torchgeo.datamodules.utils.MisconfigurationException: LEVIRCDPlusDataModule.val_dataset has length 0.

which is why the plot test fails also

You'll need to set val_split_pct=0.5 during testing, the fake data only has 2 images so 20% of 2 is 0.

[robmarkcole]

Appears an empty image is passed in the test..

torchgeo/datasets/levircd.py:240: in plot
    image1 = get_rgb(sample["image1"])
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

img = tensor([[[0., 0., 0.]]])

    def get_rgb(img: Tensor) -> "np.typing.NDArray[np.uint8]":
        img = img.permute(1, 2, 0)
        rgb_img = img.float().numpy()
        per02 = np.percentile(rgb_img, 2)
        per98 = np.percentile(rgb_img, 98)
>       rgb_img = (np.clip((rgb_img - per02) / (per98 - per02), 0, 1) * 255).astype(
            np.uint8
        )
E       RuntimeWarning: invalid value encountered in divide

[adamjstewart]

Could also create fake images with non-zero values, but dividing by epsilon is more robust.

[adamjstewart]

P.S. We actually need a ChangeDetectionTask for another project we're working on. If you have a working example, it would be good to open a PR and iterate on that.

[robmarkcole]

Re ChangeDetectionTask that's next on my list - wanted a couple of datasets to eval on first :-)

[robmarkcole]

The suggested change resulted in

Returning Any from function declared to return "ndarray[Any, dtype[unsignedinteger[_8Bit]]]"  [no-any-return]

[adamjstewart]

Yep, looks like it doesn't work

[adamjstewart]

Can you upload a new example plot to the PR description so we can make sure it looks okay?

[robmarkcole]

Updated plotting:

[robmarkcole]

Added cast back in and now get

torchgeo/datasets/levircd.py:239: error: Redundant cast to "ndarray[Any, dtype[unsignedinteger[_8Bit]]]"  [redundant-cast]

[robmarkcole]

Latest effort results in

torchgeo/datasets/levircd.py:230: error: "ndarray" expects 2 type arguments, but 1 given  [type-arg]
torchgeo/datasets/levircd.py:236: error: Returning Any from function declared to return "ndarray[Any, Any]"  [no-any-return]