I would like to train a StyleGAN2-ADA on a custom medical dataset of CT and MR images. I’m using the latest repository provided by the authors (GitHub - NVlabs/stylegan3: Official PyTorch implementation of StyleGAN3) but I see inspecting the code that the network requires images falling in the range [0 255] and with “dtype” uint8. You can see the line of code “assert image.dtype == np.uint8” in row 92 of the dataset.py script under the training folder of the repository.
I do not understand if I need to use images processed with these two constraints or if I can only comment on this line in the code
Do you have any experience with that problem?