ReMixMatch

For each unlabeled input, compute one weakly augmented current-model probability vector q. Adjust it to Normalize(q*p_prior/p_running), componentwise, using an estimated class prior and a running average of predictions on unlabeled inputs; then sharpen by a positive inverse temperature and normalize again. Hold guessed targets fixed during student differentiation. Construct one strong labeled view and K strong plus one weak view per unlabeled input, assigning its guessed target to all those views. Shuffle the combined labeled/unlabeled pair pool and use the MixMatch-style MixUp routine: a symmetric-Beta coefficient is replaced by max(lambda,1−lambda), and the same coefficient mixes input and target. Train mixed labeled and mixed unlabeled branches with cross-entropy. Also train against the guessed target on the first unmixed strong unlabeled view and apply a four-class rotation-prediction loss on that view, with the paper sampling 0,90,180,270-degree rotations uniformly. Equations3–4 write four weighted sums; the inspected author implementation instead averages each branch before applying its weights, so numerical coefficients must be interpreted with that normalization. The implementation detaches guesses, stabilizes the prior ratio with 1e-6, and uses a rolling prediction buffer. Its CTA entrypoint selects the first weak view as anchor; the standalone no-CTA entrypoint’s default swapped-view configuration is a different option. Parameter EMA is used for evaluation, not as the guessed-label teacher.