Is there sometime that sigmaloss may be inf?

[BIG-PIE-MILK-COW]

As defined in SigmaLoss, sigmaloss = -torch.log(weights) * torch.exp(-(z_vals - depths[:, None]) ** 2 / (2 * 1)) * dists, and weights = alpha * torch.cumprod(torch.cat([torch.ones((alpha.shape[0], 1)).to(alpha.device), 1. - alpha + 1e-10], -1),-1)[:, :-1], raw2alpha = lambda raw, dists, act_fn=F.relu: 1. - torch.exp(-act_fn(raw) * dists) , as sometime raw could be negative，act_fn(raw) will be 0, then raw will be 0, weights will also be 0, then log(weights) will be inf, so the sigmaloss will be inf. Then in backward, there should be some errors. Have you encountered such a situation?

[dunbar12138]

Hi, thanks for your interest. We add a small amount to weights if encountering this issue. Related issue: https://github.com/dunbar12138/DSNeRF/issues/69#issuecomment-1287234426