LitIDDPM#

class dmme.lit_modules.LitIDDPM(lr: float = 0.0002, warmup: int = 5000, decay: float = 0.9999, diffusion_model: Optional[IDDPM] = None, model: Optional[Module] = None, timesteps: int = 1000, loss_type: str = 'hybrid', gamma: float = 0.001, schedule: str = 'cosine', offset: float = 0.008, start: float = 0.0001, end: float = 0.02)[source]#

Improved Denoising Diffusion Probablistic Models

Parameters:

  • lr – learning rate, defaults to 2e-4

  • warmup – linearly increases learning rate for warmup steps until lr is reached, defaults to 5000

  • decay – EMA decay value

  • diffusion_model – overrides default diffusion_model DDPM

  • model – overrides default model passed to DDPM

  • timesteps – default timesteps passed to DDPM

  • loss_type – loss type to use either “hybrid” or “simple”

  • gamma – \(\gamma\) in hybrid loss

  • shcedule – variance schedule to use either “linear” or “cosine”

  • offset – default offset for IDDPM if cosine schedule is used

  • start – default start for IDDPM if linear schedule is used

  • end – default end for IDDPM if linear schedule is used