Repository navigation
about the partition of training and validation sets #76
Description
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?
Activity
konpatp commentedon Mar 2, 2024
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.