- CycleGAN
- StyleGAN
- StackGAN
For all existing experiments:
- Find correct hyperparameters and settings as used in papers
- Add a markdown file describing the experiment
- Commit example of generated output
- Colorization data loader (transform to luminance space etc)
- New component type: Alternator - to decide when to switch between G and D
- Make FC and Conv Generator and Discriminator more flexible
- Energy based losses
- Add a new component type to decide how to initialize layers?
- Support multiple generators/discriminators (for instance for CycleGAN)
- Find a way to add more metrics to the
lossesdict in order to write more complex or specific Evaluation classes (e.g. monitor the W1 distance) - Make sure that Tensorboard logs and snapshots can be used simultaneously
- Conditioned GAN
- Make it run under Pipenv
- Understand why we take so many CPUs (relevant: pytorch/pytorch#22866)
- Saturate GPUs
- Add documentation comments to all classes/methods
- Describe models
- Finish README with list of components
- Document the public API used for each type of component