/MattingZoom

A PyTorch-based library for trimap-free matting.

Primary LanguagePythonMIT LicenseMIT

MattingZoom

A PyTorch-based library for trimap-free matting. A complete project will be released soon!

Supported Models:

  • SHM
  • U2Net
  • MODNet
  • GFM
  • P3M-Net
  • LF
  • HATT

Supported Loaders:

  • U2Net
  • GFM
  • P3M-Net
  • FBDM

Usage

python public_worker.py --model GFM

The requirement for extending custom method.

  • model file:
    • Put the model file in the models directory.
    • The name of model file should follow [custom]_net.
    • The class name should follow [custom]_Net.
    • The parameter of init of [custom]_Net must is args.
    • The last return of forward must is the matte ~ [0,1].
  • trainer file:
    • Put the model file in the traners directory.
    • The name of traner file should follow [custom]_traner.
    • The function name should follow [custom]_Trainer.
    • The returns of [custom]_Trainer should a loss dict, which requires the first key is loss.
  • config file:
    • Put the config file in the config directory.
    • The name of config file should follow [custom]_config.yaml.
    • warming: the values in config file will become the value of args.