Skip to content

v1.5.1

Compare
Choose a tag to compare
@themattinthehatt themattinthehatt released this 31 Jul 21:13
· 29 commits to main since this release

Several small updates, as well as reverting the default training setup to automatically save out the best model according to validation metrics, rather than the last model.

  • make sure validation loader always has at least one sample
  • allow for early stopping again, as well as saving out model weights every n epochs
  • make training fully deterministic
  • increase flexibility of LitDaliWrapper to work with custom dali video pipeline functions
  • increase flexibility of keypoint correspondence array for true multiview datasets