search

angelolab / Nimbus

Other
12 stars 1 forks source link

Data aug #33

Closed JLrumberger closed 1 year ago

JLrumberger commented 1 year ago

What is the purpose of this PR?

This PR closes #26 by adding the tf augmentation pipeline and mixup.

How did you implement your changes

Added classes Flip, Rot90, GaussianNoise, GaussianBlur, Zoom, LinearContrast, MixUp and Augmenter. Augmenter applies augmentations on each image in a batch individually. MixUp works directly on batches, thus shouldn't be passed to Augmenter.

Remaining issues

None

JLrumberger commented 1 year ago

Non mixup part image Without augmentations, with augmentations and with MixUp image

In the promix paper they parameterize like this Beta(4.0,4.0), thus their actual \sigma \in [0.84, 1] which is used to weight image1 and image2 in MixUp. Honestly I don't see how MixUp improves results yet, qualitatively the images just look weird with it.