Bug in EuroSAT.plot()

[robmarkcole]

Description

I spent time trying to reproduce the EuroSAT.plot() when inferencing a single sample, and couldn't get the sample["label"] and the label shown in the plot to agree. To get them to agree it is necessary to perform sorted(EuroSAT.classes) which I do not see in the plot implementation (and doesn't make sense).

Examples below:

Steps to reproduce

# note my version without normalisation, although this shouldn't affect the plot
class EuroSATDataModule(NonGeoDataModule):
    """LightningDataModule implementation for the EuroSAT dataset.

    Uses the train/val/test splits from the dataset.

    .. versionadded:: 0.2
    """

    mean = torch.zeros(13)
    std = torch.ones(13)

    def __init__(
        self, batch_size: int = 64, num_workers: int = 0, **kwargs: Any
    ) -> None:
        """Initialize a new EuroSATDataModule instance.

        Args:
            batch_size: Size of each mini-batch.
            num_workers: Number of workers for parallel data loading.
            **kwargs: Additional keyword arguments passed to
                :class:`~torchgeo.datasets.EuroSAT`.
        """
        super().__init__(EuroSAT, batch_size, num_workers, **kwargs)


datamodule = EuroSATDataModule(
    batch_size=batch_size, 
    root="data", 
    num_workers=num_workers, 
    download=True,
)


sample = datamodule.test_dataset[0]
label = cast(int, sample["label"].item())
image = sample['image'].unsqueeze(0).to(device)
pred = task(image)
pred_index = int(torch.argmax(pred))

result_str = f"label: {sorted(EuroSAT.classes)[label]}, prediction: {EuroSAT.classes[pred_index]}" # sorted shouldn't be required!
fig = datamodule.test_dataset.plot(sample, suptitle=result_str)

Version

0.5.0

[adamjstewart]

EuroSAT subclasses NonGeoClassificationDataset which subclasses ImageFolder. ImageFolder sorts directory names to determine labels. Are the label names wrong? We can sort the class names if necessary.

[robmarkcole]

If people try to use EuroSAT.classes directly as I have, they will experience the same issue. I will make a small PR to sort these, but suggest that it is worth doing the same for other ImageFolder datasets too. I note also the difference in some label names as picked up from the folders

[robmarkcole]

Check other classification datasets:

• https://github.com/microsoft/torchgeo/blob/main/torchgeo/datasets/resisc45.py -> already sorted
• https://github.com/microsoft/torchgeo/blob/main/torchgeo/datasets/ucmerced.py -> already sorted
• https://github.com/microsoft/torchgeo/blob/main/torchgeo/datasets/patternnet.py -> no classes attribute defined

So only eurosat needs updating

[adamjstewart]

The fact that instantiating the dataset overwrites self.classes makes me kinda not want to set it at all for these datasets. We should also ensure that the order documented in the docstring matches. Thoughts?

[robmarkcole]

Agreed it doesn't need to be included, I assumed it was a convenience if you wanted to use them elsewhere. You could for example have 'pretty names' like Industrial buildings in the docstring in order, and then people could just copy from there if they need them