Are there some tips for adding custom models and dataset in the `joliGEN` lib?

[RuizhouLiu]

Thanks for your works! Are there some tips for adding custom models and dataset in the joliGEN lib?

[beniz]

Hi @RuizhouLiu see some tips below, though I guess you may have figured out most of it already:

• To add datasets, you can look at existing datareaders in data/. Some are a bit 'messy', mostly due to the internally infamous crop_image function, that does too much, but is mandatory for some of our GAN applications. Let me know if you have issues, and what you need exactly, so I can point you to the existing dataloader that is the closest to your need;

• To add a 'model', ook into the models/ dir obviously. There are two 'types' of models:

  • the img2img algorithms, such as cut, palette for diffusion, etc... Those all derive from BaseModel then from BaseDiffusionModel or BaseGANModel as needed.
  • the neural architecture for img2img, those are in models/modules. You can take the UNet with attention (https://github.com/jolibrain/joliGEN/blob/master/models/modules/unet_generator_attn/unet_generator_attn.py) as a starting example for required inputs/outputs.