nerfstudio

nerfstudio copied to clipboard

Traceback (most recent call last): File "scripts/run_train.py", line 215, in main(dcargs.cli(AnnotatedBaseConfigUnion)) File "scripts/run_train.py", line 203, in main launch( File "scripts/run_train.py", line 144, in launch main_func(local_rank=0, world_size=1, config=config) File "/home/eecs/akristoffersen/kair/pyrad/nerfactory/engine/trainer.py", line 56, in train_loop trainer.train() File "/home/eecs/akristoffersen/kair/pyrad/nerfactory/engine/trainer.py", line 142, in train self.visualizer_state.update_scene(step, self.pipeline.model, num_rays_per_batch) File "/home/eecs/akristoffersen/kair/pyrad/nerfactory/utils/decorators.py", line 58, in wrapper ret = func(self, *args, **kwargs) File "/home/eecs/akristoffersen/kair/pyrad/nerfactory/viewer/server/viewer_utils.py", line 330, in update_scene if step % num_steps == 0: ZeroDivisionError: integer division or modulo by zero

This was with running on vanilla NeRF. I think maybe the viewer did that check before any steps had occurred or something.

My guess is that the method is so slow that it can't satisfy the constraints of the fps-resolution calculation. It tries to determine how many training steps should be taken in a 1 second period such that it takes a total time of Train Util seconds. By default this value is 0.9. This logic will need to change for very slow methods ( or we present an error message in the viewer saying that it isn't capable)

Getting this issue consistently using Nerfacto on a custom polycam data set:

File "C:\Users\Rendering-PC-1\.conda\envs\nerfstudio\lib\site-packages\nerfstudio\engine\trainer.py", line 211, in train duration=self.config.pipeline.datamanager.train_num_rays_per_batch / train_t.duration, ZeroDivisionError: float division by zero

EDIT: issue still occurs with both the poster dataset and custom datasets

Fixed

I still get a ZeroDivisionError in nerfstudio\engine\trainer.py:252 with the Instant NGP implementation in nerfstudio==0.3.1 The line is duration=self.world_size * self.pipeline.datamanager.get_train_rays_per_batch() / train_t.duration where train_t.duration can become 0.0