FixMatch

Use B labeled examples and an unlabeled-input batch of size mu*B. Compute average cross-entropy on weakly augmented labeled inputs. For each unlabeled input u_b, compute q_b from the current model on its weak augmentation and make a one-hot pseudo-label at argmax(q_b). With these target predictions detached from gradients, use mask 1[max(q_b)≥tau] and cross-entropy against the model prediction on a strong augmentation of the same input. Sum masked losses and divide by the full mu*B, not by the number retained. Add this loss with fixed weight lambda_u to the labeled loss; the stated core algorithm does not require a separate weight ramp. The paper’s input pool can include labeled examples with their labels ignored in this branch. The source uses weak flip/shift augmentation and strong RandAugment or CTAugment plus Cutout; these generators and hyperparameters instantiate the procedure. EMA parameters are used for reported evaluation, while weak-view pseudo-labels come from the current model. The core does not require MixUp or distribution alignment.

Main-paper Equation4 and the inspected author code use max(q)≥tau, whereas supplementary Algorithm1 writes max(q)>tau. Equality can change whether an example contributes. The inclusive convention is used here; the supplement’s strict inequality is not silently treated as identical.