CTAugment

Maintain weights for parameter ranges of each operation, initially one, and sample a short sequence of operation types uniformly. Generate exploratory augmentations of labeled images and compare classifier probabilities with their known labels; use that feedback to update the chosen parameter weights by an exponential moving average. Use the adapted weights to select training-image augmentation parameters. This shared description is supported by the paper and released code, but their numerical conventions differ. In the pinned ReMixMatch implementation, training defaults use depth two, decay 0.99 and threshold 0.80. Probe parameters are sampled uniformly from [0,1); operation choices can repeat. The feedback score is 1−0.5*sum(abs(q−one_hot(y))), equal to the true-class probability for a valid probability vector. For training sampling, add 1−decay to the stored weights, divide by their maximum, then zero values below the threshold. Sample an adjacent parameter interval with weight equal to the sum of its two retained endpoint weights, then uniformly within that interval. Update the bin indexed by int(parameter*number_of_bins*0.999) with the observed feedback. In the ReMixMatch wrapper, the probe classifier is the returned EMA evaluation classifier. The worker generates separate probe and learned-policy training views for labeled examples, preserves the first weak view of unlabeled multiview inputs, and augments the other views with learned policies. Most parameter arrays have 17 entries, posterize has 8 and the rescale method has 6. The training wrapper overrides the bare policy class’s threshold default 0.85 with 0.80; these are implementation defaults, not universal constants.

The printed class-divided feedback score and the released-code score are different. For a valid L-class probability vector, the former is 1−(1−q_y)/L while the latter is q_y. Thus the paper expression is at least 1−1/L; interpreting its 0.8 threshold as an absolute raw-weight cutoff has different behavior from the code’s relative threshold. These are algebraic consequences of the cited formulas, not an assertion of an author-issued correction.