Training signal annealing

At training step t, compute the model probability of each labeled example’s known correct class. Keep that example in the supervised loss only when its probability does not exceed eta_t, following the inspected author implementations; hold the mask fixed during differentiation. Sum the retained cross-entropy losses and divide by max(number_retained,1), producing zero when none are retained. Set eta_t=a_t*(1−1/K)+1/K for K classes and total step count T. The source provides linear a_t=t/T, log a_t=1−exp(−5t/T), and exponential a_t=exp(5*(t/T−1)) schedules. The linear schedule spans exactly 1/K to 1 over t=0..T; the exponential begins slightly above 1/K and reaches 1, while the log schedule begins at 1/K and ends slightly below 1. These exact endpoint distinctions follow the published formulas. The threshold suppresses easy labeled examples early; this is separate from the unlabeled UDA confidence filter, which accepts high-confidence targets and averages masked losses over the full unlabeled batch.