OSCD RuntimeError: Input type (torch.cuda.LongTensor) and weight type (torch.cuda.FloatTensor) should be the same

[robmarkcole]

Description

Using the OSCD datamodule with a segmentation task results in the error:

RuntimeError: Input type (torch.cuda.LongTensor) and weight type (torch.cuda.FloatTensor) should be the same

It is necessary to cast the image to a float to resolve this (using x = batch["image"].float()), which I verified using a custom task. However I'm pretty sure this is not the intention to do this

Steps to reproduce

from lightning.pytorch import Trainer
from torchgeo.datamodules import OSCDDataModule

datamodule = OSCDDataModule(
    num_workers=num_workers,
    download=True,
    bands="rgb"
)

task = SemanticSegmentationTask(
    model="unet",
    backbone="resnet18",
    weights=True,
    in_channels=6,
    num_classes=2,
    loss="ce",
    ignore_index=None,
)

trainer = Trainer(
    min_epochs=5,
    max_epochs=25,
)

_ = trainer.fit(model=task, datamodule=datamodule)

I note that the data is type int64:

datamodule.train_dataset[0]["image"].dtype
torch.int64

Version

0.5.0

[adamjstewart]

This relates to #985. We currently don't have a good way to test this. OSCD isn't really intended for use with SemanticSegmentationTask, we need a new trainer for change detection. I would be fine with a PR that simply casts the image to float32.