IIUC, in the current implementation, the rate between original images and augmented variants will be a fixed 1 : N for N kinds of augmentation that have been enabled:
This makes it hard to control what kind of degradation to learn how fast. IMHO the precomputed files should be somehow indexed in such a way that the randomized generator can mix in augmented images (of any kind) at a well-defined rate. You would then be able to configure, say, rotation at p = 0.2. Think imgaug!
Also, I believe it should at least be possible to increase that rate throughout the training procedure.
IIUC, in the current implementation, the rate between original images and augmented variants will be a fixed
1 : N
forN
kinds of augmentation that have been enabled:https://github.com/qurator-spk/sbb_pixelwise_segmentation/blob/63fcb961898540f95cb347d43f8965ae65b8be3f/utils.py#L92
This makes it hard to control what kind of degradation to learn how fast. IMHO the precomputed files should be somehow indexed in such a way that the randomized generator can mix in augmented images (of any kind) at a well-defined rate. You would then be able to configure, say, rotation at
p = 0.2
. Thinkimgaug
!Also, I believe it should at least be possible to increase that rate throughout the training procedure.