Jupyter Notebook visualization support about torchgeo

[adamjstewart]

We could create a new LandsatTensor subclass that gets returned by the dataset, but it would get overwritten after any operation/transform, so it wouldn't help much with visualization. So it seems like it's impossible to use _repr_png_ for our use case.

The only option seems to be to do things at the Dataset level, so you ask the dataset how to plot and give it an image to plot.

[adamjstewart]

Actually, plotting a Tensor alone would be useless. What we would really like to be able to do is plot a Tensor with cartopy, which requires the crs as well. Instead of returning a Dict[str, Any], we could instead return a custom object that subclasses dict and defines how to display itself using _repr_png_.

[adamjstewart]

This custom object should still have a plot() method in addition to _repr_png_ so that it can be plotted outside of Jupyter Notebook.

[adamjstewart]

The next roadblock is how to handle samples that contain multiple images and/or labels. Specifically with labels, we need some way to define what those labels mean. I'm not sure how to do that outside of the dataset.

[adamjstewart]

We've recently added plotting functions to most datasets. I think this is sufficient for now.