Feature/16 sparse

[wronk]

Closes #16

Todo:

• Discuss logic of function
• Add tests for 3 ml-types

@drewbo, continuing earlier discussion, algorithm looks like:

1. As currently exists, get list of all tiles and labels (as one-hot vectors)
2. IF sparse flag specified, separate all tilename keys based on if they are background (i.e., >0 in the 0th position of the label) using class_match from utils.py
3. Randomly select sub-set of background images (equal to background_ratio * len(non_background_examples))
4. Continue on like normal

[drewbo]

I think the test expectations are a little off and made a comment there. Otherwise 👍

[drewbo]

We should actually expect not to see this first one, correct?

[wronk]

Deleting it didn't cause checks to fail. That was weird, but it was because we just check that every tile in the npz file matches something in expected labels. I added a line (here and in vanilla verify-labels.py) to make sure that the number of tiles in label.npz and expected_labels also match