Open mbazzani opened 1 year ago
For generating pseudo-labels, I think we should just run it without augs to get an accurate confidence value. For training on it, I think we should always data aug on stuff we're training on.
@Sean1572 @sprestrelski Thoughts on which data augs to use for finetuning?
Based on slack messages data augs seem very necessary
Should we use data augmentations/ mixup for finetuning on pseudolabels? I think data augs should be significantly less aggressive for the pseudolabeling. However, do we want that to mean a different set of augs, weaker augs, no augs, or something else entirely?