search

Mikubill / naifu

Train generative models with pytorch lightning
MIT License
284 stars 36 forks source link

[DRAFT] Dataset refactoring, support full latent cache and loss masking, separated opt for text emb training, add sparse optimizer #17

Closed AdjointOperator closed 1 year ago

AdjointOperator commented 1 year ago

Known broken staff and issues:

  1. grad acc is, for manual optimization
  2. Text emb training, for incompatible dataset class
  3. config files completely out order
  4. arb impl not identical to NAI version