Added implementation for FixMatch, modified from classification to a segmentation algorithm. code/FixMatch_standard_augs.py uses standard crops/rotations for the weak augmentations, and a color jitter for the strong augmentation. This is similar to the process implemented in PseudoSeg for segmentation. It also implements negative learning, as outlined in the CLS algorithm.
code/FixMatch_cta.py implements learned augmentations using Control Theory Augment (outlined in ReMixMatch), to learn an optimal augmentation policy during training. This model appears to give promising results, achieving >86% mean dice accuracy, based on labels from 7 patients.
Added implementation for FixMatch, modified from classification to a segmentation algorithm.
code/FixMatch_standard_augs.py
uses standard crops/rotations for the weak augmentations, and a color jitter for the strong augmentation. This is similar to the process implemented in PseudoSeg for segmentation. It also implements negative learning, as outlined in the CLS algorithm.code/FixMatch_cta.py
implements learned augmentations using Control Theory Augment (outlined in ReMixMatch), to learn an optimal augmentation policy during training. This model appears to give promising results, achieving >86% mean dice accuracy, based on labels from 7 patients.