InriaAerialImageLabelingDataModule: fix predict dimensions

[adamjstewart]

Previously, the following code would fail due to an extra dimension in predict samples:

datamodule = InriaAerialImageLabelingDataModule(...)
model = SemanticSegmentationTask(...)
trainer = Trainer(...)

trainer.fit(model=model, datamodule=datamodule)
trainer.predict(model=model, datamodule=datamodule)

This is now fixed and properly tested with a real trainer. I also removed the inria datamodule tests since they aren't useful.

[adamjstewart]

The only reason we run predict on the val set instead of the predict set is because not all datamodules have a predict set. In order to get 100% coverage, we ran everything on val instead. However, at least for segmentation, we do have a predict set, so we should use it. This is how I discovered the bug to begin with.

[adamjstewart]

@ashnair1 this likely breaks your code in #560. I can't think of an easy way to support both "b c h w" and "b n c h w" in forward, but if you can think of one let me know. Otherwise, you'll need to "b c h w -> b n c h w" manually before running CombineTensorPatches.

[ashnair1]

@ashnair1 this likely breaks your code in #560. I can't think of an easy way to support both "b c h w" and "b n c h w" in forward, but if you can think of one let me know. Otherwise, you'll need to "b c h w -> b n c h w" manually before running CombineTensorPatches.

Can't think of a way to support both. Adding the number of patches to the sample dict as suggested in my comment will allow reshaping prior to CombineTensorPatches. That should work for now.