Compute the expected output after dropout from the keep probability, plus the inference-time scaling factor that keeps train and test expectations aligned.
Dropout 率通常取 0.2-0.5。