search

autonomousvision / mip-splatting

[CVPR'24 Best Student Paper] Mip-Splatting: Alias-free 3D Gaussian Splatting
https://niujinshuchong.github.io/mip-splatting/
Other
1.07k stars 69 forks source link

Regarding training time #59

Open ShixuanGu opened 3 hours ago

ShixuanGu commented 3 hours ago

Many thanks for the great work! I'm wondering how long does it take to train a single scene?

I'm training locally on a windows machine w. 4080s, cuda 12.4, the training on "flowers", the process slows down from 9%~10%:

Computing 3D filter [01/11 16:13:43] Training progress: 7%|██▉ | 2100/30000 [00:25<08:09, 56.96it/s, Loss=0.1321220]Computing 3D filter [01/11 16:13:45] Training progress: 7%|███ | 2200/30000 [00:27<08:26, 54.89it/s, Loss=0.1205735]Computing 3D filter [01/11 16:13:47] Training progress: 8%|███▏ | 2300/30000 [00:29<08:48, 52.41it/s, Loss=0.0876850]Computing 3D filter [01/11 16:13:49] Training progress: 8%|███▎ | 2400/30000 [00:31<09:05, 50.57it/s, Loss=0.1155078]Computing 3D filter [01/11 16:13:51] Training progress: 8%|███▌ | 2500/30000 [00:33<09:32, 48.04it/s, Loss=0.1140291]Computing 3D filter [01/11 16:13:53] Training progress: 9%|███▋ | 2600/30000 [00:37<15:38, 29.20it/s, Loss=0.0955869]Computing 3D filter [01/11 16:13:58] Training progress: 9%|███▌ | 2700/30000 [03:09<10:30:46, 1.39s/it, Loss=0.1132683]Computing 3D filter [01/11 16:16:33] Training progress: 9%|███▋ | 2800/30000 [06:56<17:12:19, 2.28s/it, Loss=0.1008146]Computing 3D filter [01/11 16:20:21] Training progress: 10%|███▋ | 2860/30000 [09:39<20:47:34, 2.76s/it, Loss=0.1026932]

Is it normal?

niujinshuchong commented 2 hours ago

Hi, this is strange. The training should done in less than 1 hour. Please refer to https://github.com/autonomousvision/mip-splatting/issues/54. Could you please check how many gaussians are used?

ShixuanGu commented 1 hour ago

Many thanks for the rapid reply.

I'm training with mipnerf360, on flowers with factor = 8.

I followed the suggestions in #54 and remove selected_pts_mask = torch.logical_or(selected_pts_mask, selected_pts_mask_abs)

for the density setting, I set the threshold as 0.5: gaussians.densify_and_prune(opt.densify_grad_threshold, 0.5, scene.cameras_extent, size_threshold)

I check the Number of points at initialisation is 38347, at iteration 5100/30000, its 866647, 5200/30000 is 884200. Is that a normal setting?

ShixuanGu commented 11 minutes ago

Yet the code for nerf_synthetic works well, do you have any idea what's the potential issue?