Problem training with DataLoader

[ClaraVth]

I have isolated the training task from the .pyt script to run it without ArcGIS shutting down after the second run. In the first run, I get the error that I could reproduce in this isolated script torchgeo_logic.py.

It gives the error ValueError: A frozen dataclass was passed to 'apply_to_collection' but this is not allowed.

It should be similar to Issue #1426, but I can't implement the suggested solution.

[adamjstewart]

We have a few options we can use to solve this problem:

1. Fix this in TorchGeo (remove crs and bounds from __getitem__)
2. Fix this in PyTorch/Lightning (figure out why frozen dataclasses aren't allowed and ignore them)
3. Fix this in your code

1 would be problematic, we would like to keep this metadata if possible. 2 would be worth doing, but may take a while to get the bug fix into a release. I suggest 3 for now.

To fix this in your code, you can either:

1. Use a transform to pop the crs and bounds keys from the dataset
2. Use a GeoDataModule to pop those keys

I think 1 is easiest. Try using:

class DropFrozenKeys:
    def __call__(self, sample):
        sample.pop('crs')
        sample.pop('bounds')
        return sample

transforms = DropFrozenKeys()

image_dataset = RasterDataset(paths=image_path, transforms=transforms)
mask_dataset = RasterDataset(paths=mask_path, transforms=transforms)
mask_dataset.is_image = False

Let me know if this helps. If it does, this will be sufficient for a temporary fix. We can discuss fixing this in TorchGeo/PyTorch/Lightning more properly at a later date.

[ClaraVth]

I've reopended this issue since we need to keep the bounds for the following segmentation task.

[adamjstewart]

To keep the bounds, try something like:

bbox = sample['bounds']
sample['bounds'] = torch.tensor([bbox.minx, bbox.miny, bbox.maxx, bbox.maxy])