search

NVIDIAGameWorks / kaolin-wisp

NVIDIA Kaolin Wisp is a PyTorch library powered by NVIDIA Kaolin Core to work with neural fields (including NeRFs, NGLOD, instant-ngp and VQAD).
Other
1.45k stars 132 forks source link

Mixed precision decorators for interactive renderer #122

Closed orperel closed 1 year ago

orperel commented 1 year ago

Wisp already uses mixed precision for multiview-training, this change aligns the interactive renderer and sdf pipelines as well.

  1. @torch.cuda.amp.autocast() added to render() and redraw() functions of RendererCore and WispApp. This is toggle-able through WispState.renderer.enable_amp.
  2. BaseTrainer.iterate uses autocast on step() by default, this is controlled by the enable_amp flag passed to the trainer.

Signed-off-by: operel operel@nvidia.com