Skip to content

Conversation

@AlexandreAdam
Copy link
Owner

Change Log

  • Added EDM Score Model.
  • Added EDMv2Net.
  • Refactored some arguments in neural nets.
  • Updated default Fourier scale to 0.02 instead of 16.
  • Added Inverse square root decay for the learning rate, set to constant LR by default, in the trainer.
  • Added Karras EMA, with logic to save checkpoints according to multiple ema_lengths.
  • Added PostHoc EMA wrapper to synthesize arbitrary neural network profile from saved checkpoints, especially when multiple lengths profile are present, in which case this wrapper is used to synthesize a model at the required ema_length.
  • Included soft reset in EMA, by default is turned off.

…loading logic so that global step can be reloaded from a checkpoint
…could handle the case where multiple ema_lengths are present
…ametrization, now Karras EMA is back a default for the trainer as the preferred method.
@AlexandreAdam AlexandreAdam changed the title feat: Improving architectures feature: Improving architectures Nov 23, 2024
@AlexandreAdam AlexandreAdam merged commit 15de2ed into dev Nov 23, 2024
9 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants