POSTECH-CVLab/PyTorch-StudioGAN

Question about CONFIG_PATH in custom datasets

carlok opened this issue · 2 comments

carlok commented

Hi,
thanks for your work.
I'm trying PyTorch-StudioGAN for the first time on a custom dataset I created according to your documentation tree structure.

In the training example commands, I'm not sure how to set the CONFIG_PATH.

  1. Is it a custom yaml I should set up taking example from one of the examples in ./src/configs/?
  2. Do the images in the training/validation folders have to be squared? You didn't mention that, just wondering.
  3. Regarding image visualization, what should I pass to ckpt parameter? It should be a folder, according to main.py source but I don't understand which folder should I point it out.

Thanks.

alex4727 commented

Sorry for late reply,

  1. Correct, if you are using a custom dataset it should be pointing to your custom yaml file.
  2. No, they do not have to be squared. StudioGAN automatically performs cropping and resizing.
  3. During training, StudioGAN will create a folder like CIFAR10-ACGAN-Mod-Big-ADC-train-2022_03_10_14_23_09. ckpt should be pointing that folder.
    Thanks!
mehdiosa commented

Hi @alex4727
is this

  1. No, they do not have to be squared. StudioGAN automatically performs cropping and resizing.

also valid for StyleGAN2?

I am currently trying to get it running with a custom dataset which I load from an HDF5 File and I keep getting an error that I have the wrong dimensions:

AssertionError: Wrong size for dimension 3: got 64, expected 256
  • 👀2