nerfacc rendering gradient error

[SYSUykLin]

nerfacc rendering gradient error

[dragonwarrior]

i had a similar problem that 'nerfacc.rendering' can't bp gradients, why

here is my code:

        def _rgb_alpha_fn(t_starts, t_ends, ray_indices):
            ray_indices = ray_indices.long()
            t_origins = rays_o[ray_indices]
            t_dirs = rays_d[ray_indices]
            midpoints = t_starts + (t_ends - t_starts) * torch.rand_like(t_starts)

            positions = t_origins + t_dirs * midpoints[..., None]
            dists = t_ends - t_starts
            
            sdf_nn_output = self.sdf_network(positions)
            sdf = sdf_nn_output[:, :1]
            feature_vector = sdf_nn_output[:, 1:]

            sdf_grad = self.sdf_network.gradient(positions).squeeze()
            rgbs = self.color_network(positions, sdf_grad, t_dirs, feature_vector).reshape(-1, 3)

            normal = F.normalize(sdf_grad, p=2, dim=-1)
            alphas = self.get_alpha(sdf, normal, t_dirs, dists, cos_anneal_ratio)

            return rgbs, alphas

        comp_rgb, opacity, depth, extras = nerfacc.rendering(t_starts, t_ends, ray_indices, n_rays=n_rays, rgb_alpha_fn=_rgb_alpha_fn)
        fake_loss = comp_rgb.sum()
        fake_loss.backward()
        print(self.sdf_network.lin0.weight.grad) # no grad here, why???

[SYSUykLin]

do u use the estimater.sample? I set the alpha_thre and got zero ray_indeces, I delete the alpha_thre in estimater.sample and gradient present.

[dragonwarrior]

yes, i used estimator.sample api.
right now, i solved my problem, it's because another mistake made by myself. the api is ok.