the real data in **train.lst**(more than 48k) or **train_pairs.txt**(more than 37k) , which one is used to calculate the metric, FID?