Skip to content

Latest commit

 

History

History
61 lines (46 loc) · 2.07 KB

flow.md

File metadata and controls

61 lines (46 loc) · 2.07 KB

DL_diffusion_model flow

Dataset

  • MNIST
  • Fashions mnist?
  • celebA?
  • Cifar10
  • Vettori numerici

Trainer

  • Trainstep method
  • Train method
  • Test train method
  • Performance Meter
  • Model Save / Load
  • Check Point / Tensorboard
  • Evaluation method / Evaluator Object (maybe not necessary)

DiffusionModel

UNet o stato dell'arte CNNs x image segmentation

Examples:

1. Small(no time embedding): https://github.com/dbasso98/GANs/blob/main/CycleGANs/generator.py
2. Large(with time embedding): https://colab.research.google.com/drive/1NFxjNI-UIR7Ku0KERmv7Yb_586vHQW43?usp=sharing#scrollTo=fc4IdL5YawTN
3. Simple(especially time embedding): https://huggingface.co/blog/annotated-diffusion

Loss

  • MSE noise: $L_{simple}$
  • Add $L_{VLB}$: $L_{simple}+\lambda L_{VLB}$
  • Cosine Schedule
  • Add covariance matrix
  • Improve gradient noise

Ref

Possible Future Implementations

  • DDIM