Why FPS is so low?

[chenj02]

I reproduce the train and render code, but the FPS is so low. Why?

root@60b4fd61d0f0:/data/chenj02/3DGS/code/4DGaussians# python render.py --model_path "/data/chenj02/3DGS/train_model_4dgs/banana/"  --skip_train --configs arguments/hypernerf/banana.py
Looking for config file in /data/chenj02/3DGS/train_model_4dgs/banana/cfg_args
Config file found: /data/chenj02/3DGS/train_model_4dgs/banana/cfg_args
Rendering  /data/chenj02/3DGS/train_model_4dgs/banana/
feature_dim: 48 [24/10 13:54:38]
Loading trained model at iteration 14000 [24/10 13:54:38]
load finished [24/10 13:54:40]
83it [00:00, 30895.21it/s]
format finished [24/10 13:54:40]
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 83/83 [00:19<00:00,  4.37it/s]
Loading Training Cameras [24/10 13:55:00]
Loading Test Cameras [24/10 13:55:00]
Loading Video Cameras [24/10 13:55:00]
Deformation Net Set aabb [45.55028915 10.94474697 72.02857971] [-10.79911232 -16.84693146   6.11368752] [24/10 13:55:00]
Voxel Plane: set aabb= Parameter containing:
tensor([[ 45.5503,  10.9447,  72.0286],
        [-10.7991, -16.8469,   6.1137]]) [24/10 13:55:00]
loading model from exists/data/chenj02/3DGS/train_model_4dgs/banana/point_cloud/iteration_14000 [24/10 13:55:00]
point nums: 113343 [24/10 13:55:01]
Rendering progress: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 82/82 [00:27<00:00,  2.95it/s]
FPS: 2.9202738542724496 [24/10 13:55:28]
IMAGEIO FFMPEG_WRITER WARNING: input image is not divisible by macro_block_size=16, resizing from (536, 960) to (544, 960) to ensure video compatibility with most codecs and players. To prevent resizing, make your input image divisible by the macro_block_size or set the macro_block_size to 1 (risking incompatibility).
[swscaler @ 0x66c9d00] Warning: data is not aligned! This can lead to a speed loss
point nums: 113343 [24/10 13:56:15]
Rendering progress: 330it [01:56,  2.84it/s]
FPS: 0.697885301603404 [24/10 13:58:11]
IMAGEIO FFMPEG_WRITER WARNING: input image is not divisible by macro_block_size=16, resizing from (536, 960) to (544, 960) to ensure video compatibility with most codecs and players. To prevent resizing, make your input image divisible by the macro_block_size or set the macro_block_size to 1 (risking incompatibility).
[swscaler @ 0x6544d00] Warning: data is not aligned! This can lead to a speed loss

[guanjunwu]

I dont think it's normal.
What GPU are u use, and maybe there are any scripts running on your machine?
could you test another scene such as D-NeRF?

[chenj02]

I dont think it's normal. What GPU are u use, and maybe there are any scripts running on your machine? could you test another scene such as D-NeRF?

Hi, I use a single 3090 GPU, and I also try to render in dnerf/bouncingballs, but it's also very low. The PSNR and MS-SSIM are normal, only FPS is strange. Is the reason that I use docker?

python render.py --model_path "/data/chenj02/3DGS/train_model_4dgs/dnerf/bouncingballs/" --skip_train --configs arguments/dnerf/bouncingballs.py

Looking for config file in /data/chenj02/3DGS/train_model_4dgs/dnerf/bouncingballs/cfg_args
Config file found: /data/chenj02/3DGS/train_model_4dgs/dnerf/bouncingballs/cfg_args
Rendering  /data/chenj02/3DGS/train_model_4dgs/dnerf/bouncingballs/
feature_dim: 64 [27/10 05:36:51]
Loading trained model at iteration 20000 [27/10 05:36:51]
Found transforms_train.json file, assuming Blender data set! [27/10 05:36:51]
Reading Training Transforms [27/10 05:36:51]
Reading Test Transforms [27/10 05:38:45]
Generating Video Transforms [27/10 05:38:51]
hello!!!! [27/10 05:38:51]
Generating random point cloud (2000)... [27/10 05:38:52]
Loading Training Cameras [27/10 05:38:52]
Loading Test Cameras [27/10 05:38:52]
Loading Video Cameras [27/10 05:38:52]
Deformation Net Set aabb [1.29982098 1.29990645 1.29988719] [-1.29980838 -1.29981163 -1.29872349] [27/10 05:38:52]
Voxel Plane: set aabb= Parameter containing:
tensor([[ 1.2998,  1.2999,  1.2999],
        [-1.2998, -1.2998, -1.2987]]) [27/10 05:38:52]
loading model from exists/data/chenj02/3DGS/train_model_4dgs/dnerf/bouncingballs/point_cloud/iteration_20000 [27/10 05:38:52]
point nums: 27419 [27/10 05:38:52]
Rendering progress:   0% 0/20 [00:00<?, ?it/s]
Rendering progress: 100% 20/20 [00:02<00:00,  7.38it/s]
FPS: 7.063135526221561 [27/10 05:38:55]
point nums: 27419 [27/10 05:39:06]
Rendering progress:   0% 0/160 [00:00<?, ?it/s]
Rendering progress: 100% 160/160 [00:17<00:00,  9.40it/s]
FPS: 9.340526556943423 [27/10 05:39:23]

[guanjunwu]

maybe yes, I also run the code on the 3090.