Open
Description
I am trying to implement a depth loss, in nerf/utils.py/Trainer.train_step.
outputs = self.model.render(rays_o, rays_d, staged=False, bg_color=bg_color, perturb=True, force_all_rays=False if self.opt.patch_size == 1 else True, **vars(self.opt))
the outputs['depth'] always return 0 or nan during training.
depth = outputs['depth']
print(depth[~depth.isnan()].sum()) --> always 0.
With self.model.eval(), I get an appropriate depth result with some non zero values, but unable to backprop.
Any idea what is going on?
Metadata
Metadata
Assignees
Labels
No labels