Hi, could you explain the configuration when calculating the Fréchet Inception Distance? Specifically:
- Which dataset was used to calculate the mean and covariance of real world samples?All training samples(train, val and test set)? Only train set samples? Or only test set samples?
- The details of calculating the mean and covariance of generated samples? Did it follow the sg2im paper with five samples generated for each image on COCO test set?
Thanks.