UCMerced: fix image shape bug

[isaaccorley]

Previously we performed some preprocessing transforms for each individual sample. In #992 we changed this to perform all transforms on a batch.

However, some of the UCMerced images vary in size, e.g. 247, 254, 248. This causes an issue when loading batches of the imagery because they aren't the same shape and aren't being resized. This PR adds a default resize transform so that if a user doesn't specify the transform as a kwarg, the dataset will still work with a dataloader.

[calebrob6]

literally just ran into this

However, some of the UCMerced images vary in size, e.g. 247, 254, 248.
253

[adamjstewart]

Can you change the fake data for UC Merced to reflect the fact that the image size isn't consistent? This will let us test things properly.

[isaaccorley]

Something seems to be wrong with the tests because the test data does vary in size already

agricultural02.tif (247, 247, 3)
agricultural00.tif (256, 256, 3)
agricultural01.tif (249, 256, 3)

[isaaccorley]

Ok I figured it out. It's because in the datamodule inside the test config we are using a batch size = 1, so it doesn't have to combine 2 images of different shapes into a batch unless batch_size > 2

[adamjstewart]

Is it worth resizing in the dataset instead of the datamodule?

[isaaccorley]

Is it worth resizing in the dataset instead of the datamodule?

Good point. I think keeping this in the datamodule is better because it doesn't make any assumptions about how the user wants to modify the dataset.

[adamjstewart]

I guess the benefit of doing it in the dataset is that users who use the dataset without a datamodule won't also have to do this. Are all the images approximately the same size? Looking at our existing code, we have 2 datasets (cyclone, seco) and 0 datamodules that do resizing, unless I miscounted.

[isaaccorley]

Yes they are all within a few pixels from 256x256. I'm good either way so I'll leave it up to you, just let me know so I can implement it.

[calebrob6]

I like it in the dataset so users don't have to fool around with it. Without it someone (me) is going to do:

ds = UCMerced(...)
dl = DataLoader(ds, ...)

for batch in dl:
    ...

then get the error about size mismatch, groan, go lookup how to do resize and pass that as a transform, and continue with whatever they were doing.

updated

[isaaccorley]

Something wrong with the pyvista docs

[adamjstewart]

Need to rebase now that we are using Python 3.9+ type hints