about the partition of training and validation sets

[string-ellipses]

I am confused about the division of training and validation sets, because I see "self.val_data = self.train_data ” in experiment.py (line 184), which seems to cause the training set and validation set to not be separatedd, thus affecting a fair evaluation of model performance.

perhaps there are steps in other sections to separate the training and validation sets?

[konpatp]

Validation in generative models is done a bit differently. For example, FID calculation takes 50k images to get a reasonably stable result. If all of the 50k images are set aside for validation, there will be only 20k images left for training which is impractical.