I was going through the paper and they are training logistic regression model on the code itself, not on the output of the Generator. But, in your lr code, you are training logistics regression model on the generator output. Can you explain the reason, why?