Selecting Data Augmentations

[jonwomack]

I don't see any transformations/augmentations listed in the CDFM paper, however ZooplanktoNet achieves higher classification with performance with the following augmentations:

• rotation: random with angle from 0◦ to 360◦ with 90◦
increments
• translation: random with shift −40 or 40 piexls in
horizontal and vertical direction
• rescaling: random with scale factor between 1/1.2 and
1.2
• shearing: random with angle between −20◦ and 20◦
• flipping: random in horizontal direction or vertical direction

We could cover all of these except flipping with torchvision.transforms.RandomAffine.