Apply inverted dropout from a supplied binary mask during training and preserve the input unchanged during evaluation.
